Paper deep dive
Work2Skill: Capturing Human GUI Experience for Reusable Agent Skill Induction
Yuqi Huang, Hui Li, Yifei Cao, Zikun Cai, Tianci Xie, Yangfan Gao, Pinxue Guo, Chaoyun Zhang, Bingxuan Li, Rui Zheng, Hang Yan, Ning Liao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/12/2026, 6:42:10 AM
Summary
The paper introduces Work2Skill, a framework for capturing human GUI experience to induce reusable agent skills. It comprises WorkRecorder for collecting synchronized multimodal evidence (screen, mouse/keyboard, accessibility tree), WorkDemo-350, a dataset of 350 human demonstrations, and an agentic skill-induction method that inspects evidence on demand. The framework significantly improves performance on MyPCBench and OSWorld-Verified benchmarks compared to baselines.
Entities (8)
Relation Signals (7)
Work2Skill â includes â WorkRecorder
confidence 95% ¡ WorkRecorder provides the framework's infrastructure for collecting this experience
Work2Skill â includes â WorkDemo-350
confidence 95% ¡ Using WorkRecorder, we construct WorkDemo-350... We thus introduce Work2Skill
WorkRecorder â captures â Accessibility Tree
confidence 90% ¡ aligning screen recordings, mouse and keyboard events, and available accessibility-tree information.
WorkDemo-350 â covers â MyPCBench
confidence 90% ¡ It contains 350 human GUI demonstrations: 64 MyPCBench Bounded Action tasks
WorkDemo-350 â covers â OSWorld-Verified
confidence 90% ¡ and a fixed 286-task OSWorld-Verified subset.
Work2Skill â outperforms â ShowUI-Aloha
confidence 90% ¡ Work2Skill achieves the highest scores and success rates... compare Work2Skill with... ShowUI-Aloha
Work2Skill â outperforms â Demo2Tutorial
confidence 90% ¡ Work2Skill achieves the highest scores and success rates... compare Work2Skill with... Demo2Tutorial
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Computer-use agents can perform increasingly complex GUI tasks, yet unfamiliar software and workflows requiring human domain knowledge remain challenging. Human demonstrations convey procedural knowledge through the same GUI interaction space, but reuse requires interpreting actions in context and separating workflow rules from task-specific details. We thus introduce Work2Skill, a framework that captures human GUI experience for reusable agent skill induction. WorkRecorder provides the framework's infrastructure for collecting this experience, aligning screen recordings, mouse and keyboard events, and available accessibility-tree information. Using WorkRecorder, we construct WorkDemo-350, the first synchronized multimodal dataset of human desktop GUI demonstrations for reusable agent skill induction and downstream evaluation. Its 350 demonstrations are screened for completeness and task success. We then propose an agentic skill-induction method that inspects evidence on demand to produce reusable skills capturing workflow rules, step dependencies, and completion checks. When the same four execution models are equipped with skills induced by each method under shared execution settings, Work2Skill achieves the highest scores and success rates. Mean score improvements over the corresponding No Skill baseline are 4.67--10.54 points on MyPCBench Bounded Action and 3.11--7.50 points on a 286-task subset of OSWorld-Verified. Our code and data are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.12984v1
- Canonical: https://arxiv.org/abs/2610.12984v1
Trouble viewing inline? Open PDF directly â
Full Text
128,529 characters extracted from source content.
Expand or collapse full text
Preprint WORK2SKILL: CAPTURING HUMAN GUI EXPERI- ENCE FOR REUSABLE AGENT SKILL INDUCTION Yuqi Huang 1,⥠Hui Li 2,⥠Yifei Cao 2 Zikun Cai 3 Tianci Xie 4 Yangfan Gao 2 Pinxue Guo 5,â Chaoyun Zhang 5,â Bingxuan Li 5 Rui Zheng 5 Hang Yan 5 Ning Liao 1,â 1 Shanghai Jiao Tong University 2 Fudan University 3 Nanyang Technological University 4 Zhejiang University 5 Shanghai Qiji Zhifeng Co., Ltd ⥠Equal contribution â Project Leaders * Corresponding Author ABSTRACT Computer-use agents can perform increasingly complex GUI tasks, yet unfamil- iar software and workflows requiring human domain knowledge remain challeng- ing. Human demonstrations convey procedural knowledge through the same GUI interaction space, but reuse requires interpreting actions in context and separat- ing workflow rules from task-specific details. We thus introduce Work2Skill, a framework that captures human GUI experience for reusable agent skill induc- tion. WorkRecorder provides the frameworkâs infrastructure for collecting this experience, aligning screen recordings, mouse and keyboard events, and available accessibility-tree information. Using WorkRecorder, we construct WorkDemo- 350, the first synchronized multimodal dataset of human desktop GUI demon- strations for reusable agent skill induction and downstream evaluation. Its 350 demonstrations are screened for completeness and task success. We then pro- pose an agentic skill-induction method that inspects evidence on demand to pro- duce reusable skills capturing workflow rules, step dependencies, and completion checks. When the same four execution models are equipped with skills induced by each method under shared execution settings, Work2Skill achieves the highest scores and success rates. Mean score improvements over the corresponding No Skill baseline are 4.67â10.54 points on MyPCBench Bounded Action and 3.11â 7.50 points on a 286-task subset of OSWorld-Verified. Our code and data are available at https://github.com/china-qijizhifeng/Work2Skill. 1INTRODUCTION Work2Skill WorkRecorder Capturehuman GUI experience WorkDemo-350 350 curated demonstrations Agentic Skill Induction Induce reusable agent skills Figure 1: Work2Skill framework: WorkRecorder, WorkDemo-350, and Agentic Skill Induction. Computer-use agents can perform increasingly complex desktop and web tasks, supported by advances in GUI grounding and plan- ning (Qin et al., 2025; Agashe et al., 2025; Wang et al., 2025a). Yet agents still struggle with unfamiliar software and workflows that require domain knowledge, where successful execution depends on procedures not spelled out in the userâs request (Zhang et al., 2026e; Jang et al., 2026). For example, âsubmit my reimburse- mentâ does not specify which internal portal to use, which receipts to attach, or what to check before submission. Agents need this user- or organization-specific procedural knowledge to translate a userâs goal into GUI actions. Human demonstrations offer a direct source of this knowledge be- cause humans and computer-use agents share the same GUI in- teraction space.Demonstrated controls, action sequences, and visible outcomes can therefore guide an agentâs own execu- tion.Prior work converts demonstrations into expert instruc- tions and semantic action traces (Li et al., 2025; Zhang et al., 2026e), multimodal tutorials (Bai et al., 2026), or reusable skills 1 arXiv:2610.12984v1 [cs.AI] 9 Oct 2026 Preprint (Yang et al., 2026a; Xiao et al., 2026). OpenAIâs Record & Replay (OpenAI, 2026b) and Claudeâs Record a skill (Anthropic, 2026b) bring this approach to user-facing tools. We thus introduce Work2Skill, a framework that captures human GUI experience for reusable agent skill induction (Figure 1). Its three components address the following challenges in action grounding, method comparison, and evidence inspection. First, action coordinates reveal where an operation occurred but not what the target control means. ShowUI-Aloha interprets actions from full-screen frames and target crops (Zhang et al., 2026e), while Demo2Tutorial uses before-and-after screenshots (Bai et al., 2026). Without accessibility- tree (AX) evidence, these parsers depend on visual grounding to infer target semantics; selected frames can also miss transient states and intermediate operations. WorkRecorder addresses this ambiguity by aligning screen recordings, mouse and keyboard events, and available AX information. Recorded element names and roles provide explicit control semantics, while linked video preserves the surrounding state changes. The induction agent can thus interpret an operation using both its visual context and available structured UI evidence. Second, representative desktop GUI studies use different demonstrations, task sets, and execution protocols (Zhang et al., 2026e; Bai et al., 2026). Their results mix differences in induction with differences in source experience and evaluation conditions. WorkDemo-350 provides a common demonstration resource for skill induction in desktop environments. Collected with WorkRecorder, it contains 350 human GUI demonstrations covering 64 MyPCBench Bounded Action tasks (Jang et al., 2026) and a 286-task OSWorld-Verified subset (Xie et al., 2024). Each demonstration is screened for completeness and task success and packaged with synchronized visual and action evi- dence, plus available AX information. This shared evidence and a controlled downstream protocol support comparisons of induction methods using the same task evaluators and execution conditions. Third, representative pipelines prescribe an order for action parsing and procedure abstraction, even when individual stages refine their outputs (Zhang et al., 2026e; Bai et al., 2026). This couples ev- idence inspection to predefined stages. Loading a long recording into context at once also creates context pressure, while selecting a fixed subset can omit details needed later. Our agentic skill- induction method addresses both constraints by interleaving on-demand evidence inspection with skill writing and revision. Drawing on active evidence acquisition in agentic video understand- ing (Yuan et al., 2025; Wang et al., 2026; Li et al., 2026), the agent selects linked files, images, and video intervals to resolve questions as they arise. It can revisit the source after drafting with- out loading the entire recording at once. During skill writing, we guide the agent to generalize demonstration-specific inputs into task-dependent descriptions while preserving procedural rules, step dependencies, and completion checks. The skills guide execution without the source recording. We evaluate the resulting skills through downstream execution on the WorkDemo-350 tasks. Us- ing official evaluators and a shared text-guidance interface, we compare Work2Skill with No Skill and text-adapted versions of ShowUI-Aloha and Demo2Tutorial. Across four execution models, Work2Skill improves mean evaluation scores over No Skill by 4.67â10.54 points on MyPCBench Bounded Action and 3.11â7.50 points on the evaluated OSWorld-Verified subset. Further experi- ments test reuse under changed requirements within the same application (MyPCBench Jitter) and transfer between two execution models. Evidence ablations and the Pipeline Prompt comparison examine the roles of input modalities and induction-prompt design. Our contributions are threefold: 1) We introduce WorkRecorder, the human GUI experience collection infrastructure of Work2Skill, which aligns recorded inputs with visual and available AX evidence to clarify demonstrated actions. 2) We construct WorkDemo-350, the first synchronized multimodal dataset of human desktop GUI demonstrations for reusable agent skill induction and downstream evaluation. 3) We propose an agentic skill induction method that interleaves on-demand evidence inspection with skill writing and revision. Across two benchmarks and four execution models, it achieves the highest scores and success rates among the evaluated methods. 2RELATED WORK Learning from Human GUI Demonstrations. Human demonstrations provide experience for GUI agents across desktop, web, and mobile environments (Nguyen et al., 2025; Tang et al., 2025; 2 Preprint Task Instruction Task: Update March Food Budget to 450 and Save. Online Environments Benchmark Environments (e.g. MyPCBench / OSWorld) File System Web Database Terminal Text Editor Settings AppEmail ...... #4 Agentic Skill Induction Video Keyframe Images Input Events Accessibility EvidenceStates (APP/OS) Action-triggered keyframe capture Timestamp-based evidence association Recording export & benchmark conversion Recording Properties Click[x,y]Move[x,y]<div>...<div>App:Excel Build Evidence Chain(per step) Step t 0:10:20 Step t+1 0:11:15 Step T 0:18:04 time Screenshot Action: Click Input: None Accessibility : Cell Selected Screenshot Action: Type Input: â450â Accessibility: Cell Value=450 Screenshot Action: Keyboard Shortcut Input: Ctrl+S Accessibility : Current Application Overall Performance(GPT-5.5) Evaluation Workflow Induced Skill MyPCBench OSWorld-286 +7.50 score points Compared to No Skill +10.54 score points Compared to No Skill Benchmarks ¡ MyPCBench ¡ OSWorld-286 Environments Baselines ¡ No skill ¡ ShowUI-Aloha ¡ Demo2Tutorial Skill extraction VM execution Analysis ... Adaptive exploration outperformsfixed evidence processing. Multimodal evidence improvesskill quality. Key Analyses & Findings Induced skills generalize across models and tasks. 1 2 3 Set Up Human Operations #5 Execution-based Evaluation #3Action-linked Evidence Construction #1 Human GUI Experience #2 Multimodal Experience Recording 3.Update Hypothesis & Resolve Uncertainty 2.Explore with Tools1.Identify Information Needs 4.Skill Workspace What? How? Which? Task: Update March Food Budget to 450 and Save. Current Evidence: â˘Target application â˘Object identifier â˘Input method â˘...... Evidence Exploration Tools list_directoryglobread_visual _evidence read_file watch_ video extract_video_ frames get_video_ metadata finalize_extr action Cross-modal verification <cell name=âB3 ... Refined hypothesis Open budget.xlsx Select cell B3 Relevant multimodal evidence reviewed Task goals and constraints preserved Reusable skill finalized for execution WorkRecorder Workdemo-350 Figure 2: Overview of Work2Skill. WorkRecorder aligns human GUI actions with visual and avail- able accessibility evidence to build WorkDemo-350. An agent inspects evidence on demand and revises reusable skills for downstream GUI execution. Downstream evaluation compares these skills with guidance from the fixed workflows of ShowUI-Aloha and Demo2Tutorial. Qin et al., 2025). They support data collection, grounding, model training (Wang et al., 2025a; Nayak et al., 2025; Feizi et al., 2026) and agent efficiency evaluation (Abhyankar et al., 2026). Demonstration structure can also influence task execution (Xing et al., 2026). Procedural represen- tations include expert instructions (Li et al., 2025), task graphs (Yin et al., 2025), and parameterized workflows for robust replay (Zhao et al., 2026). LearnAct also introduces LearnGUI for evaluat- ing demonstration-based learning on mobile devices (Liu et al., 2026). More recent work includes semantic action traces, multimodal tutorials, natural-language skill graphs, and standard operating procedure (SOP) skills (Zhang et al., 2026e; Bai et al., 2026; Yang et al., 2026a; Xiao et al., 2026). Skill Induction and Active Evidence Acquisition. Reusable guidance draws on experience or- ganization and evidence selection. Skill-induction methods learn procedures through autonomous interaction (Wang et al., 2025b; Zheng et al., 2025) or distill execution and user traces (Mi et al., 2026; Ni et al., 2026; Yang et al., 2026c), with environment feedback and policy learning supporting further refinement (Xia et al., 2026; Yang et al., 2026b). Adaptive memories, contextual playbooks, and learned memory-management routines retain experience (Suzgun et al., 2026; Zhang et al., 2026c;a), while experience-based evaluation examines its downstream value (Peng et al., 2026). Guidance from multimodal demonstrations requires choosing what to inspect. Agentic video meth- ods retrieve and examine relevant segments iteratively (Ma et al., 2025; Yang et al., 2025; Yuan et al., 2025; Wang et al., 2026; Li et al., 2026). In GUI settings, CUADebug inspects traces and screenshots to diagnose execution failures (Zhang et al., 2026d), while MMSkills packages textual procedures, state cards, and keyframes for selective inspection during execution (Zhang et al., 2026b). 3WORK2SKILL FRAMEWORK Work2Skill connects human demonstrations to reusable procedural guidance (Figure 2). WorkRecorder preserves actions together with their visual and semantic context; WorkDemo-350 organizes the collected demonstrations into a curated evidence resource. An induction agent in- spects this evidence, revisits unclear steps, and writes skills for downstream GUI execution. 3.1WORKRECORDER WorkRecorder collects the evidence needed to interpret human actions in context. Given a task instruction, a human demonstrator performs the task in the benchmarkâs official environment. 3 Preprint Recorder Interface Packaged Evidence Episode Multimodal Experience Recording Trace/ Validated and packaged Aligned timestamps Ready for downstream use cua_only-f002/ Evidence/ 000001.png 000002.png Recording Status Environment QEMU-TigerVNC Example Case âTASKâ: âIstartedthisyearâstaxreturnonSpeedTax butneverfinished........ Don'tsubmitityet,justgetit readyformetoeyeball." Environment Response Left Click combo box -Search with Google or enter address Human Action 0:02:04 Left Click menu item -SpeedTax Human Action 0:02:58 Scroll stopped (14 events merged) Human Action 0:03:17 Left Click push button -Review again Human Action 0:16:54 Left Click link -Back to overview Human Action 0:17:09 ...... ...... task.json episode.json video.mp4 summary.md actions.jsonl keylog.jsonl image_actions.jsonl manifest.json â˘Capture video, screenshots, mouse & keyboard, semantic actions, UI element, APP state. â˘All streams recorded and synchronized 1 2 3 4 Task ID cua_only-f002 Duration 00:03:24(in progress) State Recording...... Environment Response Environment Response Environment Response Environment Response Figure 3: Human GUI experience collection with WorkRecorder. A demonstrated workflow (left) is captured and reviewed through the recorder interface (center), then organized into an evidence episode with task metadata, action traces, and available multimodal evidence (right). WorkRecorder captures screen video, keyframe images, mouse and keyboard activity, and available UI information as multimodal evidence (Figure 3). An action trace links each recorded operation to its supporting evidence so that the induction agent can inspect individual steps. Task setup and outcome evaluation follow the corresponding benchmark procedures. A macOS interface records local workflows or interactions with a remote Linux desktop. For remote tasks, the Mac captures the displayed workflow and local inputs, while a collector in the target environment supplies UI evidence. Platform-specific collectors provide target-element metadata, accessibility-tree (AX) snapshots, and application or window context when available. Coverage depends on the recording configuration and the applicationâs accessibility support. Multimodal Capture and Alignment.WorkRecorder aligns input events and visual observations on a common video-relative timeline. It groups timestamped events into actions such as clicks, drags, scrolling, and hotkeys, then links these actions to keyframes. An agent can therefore inspect a click alongside its screen image, available target information, and surrounding video to determine what changed. Target-element association uses pointer location and event timing, with window context and control semantics helping select among candidates. On Linux, AT-SPI point queries and nearby accessibility events supply available names, roles, and bounds for click and drag records. UI annotations are derived from the corresponding target elements in the AX tree. AX snapshots provide additional interface context but need not reflect the exact action-time state. Appendix A details the recording architecture, capture policies, target-matching rules, and time alignment. Evidence Representation and Packaging.Each demonstration becomes an evidence episode containing the task instruction and metadata, an ordered action trace, and the available multimodal evidence. A manifest indexes these resources; action records link to keyframes and video times. These links support targeted inspection without processing the recording sequentially. Descriptions help locate operations; linked recordings let the agent verify their interpretation. The available evidence depends on capture and export settings. Appendix B provides details. 3.2WORKDEMO-350 WorkDemo-350 provides a common evidence resource for studying skill induction. Collected with WorkRecorder, it contains 350 human GUI demonstrations: 64 MyPCBench Bounded Action tasks and a fixed 286-task OSWorld-Verified subset. The two benchmarks provide complementary set- tings. MyPCBench (Jang et al., 2026) covers personal office and everyday workflows with user- specific context; its Bounded Action tasks offer scoped GUI workflows suited to human demonstra- 4 Preprint tion. OSWorld-Verified (Xie et al., 2024) adds general computer-use tasks across desktop and web applications, broadening the range of workflows available for skill induction and evaluation. Table 1: Statistics of WorkDemo-350: 350 human GUI demonstrations covering 64 MyPCBench Bounded Ac- tion tasks and a 286-task subset of OSWorld-Verified, with aligned multimodal evidence. StatisticMyPCBench OSWorld-Verified Number of tasks64286 Total video duration (s)7623.0222890.18 Mean episode duration (s)119.1180.04 Mean actions per episode32.2334.57 Mean evidence images24.2023.91 We screen tasks and recordings to re- tain successful, high-quality demon- strations. For OSWorld-Verified, we first exclude infeasible tasks for which a valid human demonstration cannot be completed. Across both benchmarks, demonstrators follow valid GUI work- flows and verify the final state without being restricted to a prescribed action sequence. We then review task suc- cess and recording completeness, dis- carding demonstrations that omit key operations or end in an incorrect state (Appendix E). Each episode links task and action records to available visual and UI evidence. Episodes are organized by benchmark and task so that methods can draw on the same demonstra- tions while choosing which modalities and time intervals to inspect. The collection contains ap- proximately 8.48 hours of recorded experience, accompanied by action traces and evidence images. Table 1 summarizes the datasetâs size and multimodal evidence statistics. 3.3AGENTIC SKILL INDUCTION Skill induction turns an evidence episode into reusable procedural guidance. Given the task instruc- tion and recorded evidence, a NexAU tool-using agent (Nex-AGI, 2026) generates skills in an output directory separate from the recordings. The default system prompt is concise: it sets evidence-access boundaries, prerequisite checks, and completion requirements while leaving the detailed exploration and revision sequence to the agent. A bundled skill-creator policy guides skill authoring. After reading this policy and the task definition in task.json, the agent alternates evidence inspection with skill writing and revision. Appendix D provides the prompts and authoring policy. Evidence Exploration Tools.We select file, image, and video tools that expose complemen- tary views of a demonstration at different granularities. File discovery and reading provide task definitions, manifests, action traces, key logs, and available structured UI evidence. Traces and in- dices link operations to their supporting media, while AX records supply available control names and roles. read visualevidence returns selected images with source-path labels, exposing visible controls and entered or selected values. Video tools provide access to states and transi- tions beyond the indexed images. After reading video metadata, the agent can sample an interval with watchvideo or request individual timestamps with extractvideoframes; both return timestamp-labeled frames. Source-path and timestamp labels preserve the correspondence between returned observations and source artifacts. Appendix C details the interfaces, and Section 4.5 eval- uates the contribution of the available evidence to downstream execution. Active Evidence Acquisition.The agent chooses its next observation according to the task and unresolved questions in the evidence, following iterative inspection in agentic video understanding (Yang et al., 2025; Wang et al., 2026; Li et al., 2026). The prompt requires representative initial, crit- ical, and final-state images to be inspected when indexed images are available; the agent otherwise determines the inspection sequence and level of detail. For example, an ambiguous click can prompt inspection of the target frame to identify the control, followed by a later video frame to check the resulting state. Tool guidance encourages coarse video sampling followed by closer examination of relevant actions. Evidence inspection and skill writing share one conversation, so a draft can ex- pose missing constraints or unverified transitions. The agent can then return to the relevant evidence and revise the procedure to incorporate supported dependencies and completion checks. Candidate skills are not executed during induction, and downstream evaluation results are not used as revision feedback. When the available observations do not resolve an uncertainty, the authoring policy re- quires an observable conditional check or an explicit limitation in the skill. Section 4.5 compares this prompt with Pipeline Prompt, which sets exploration order but allows further inspection. 5 Preprint Skill Generation.The agent expresses each coherent reusable capability in a SKILL.md, with YAML frontmatter containing its name and description, following the Agent Skills file convention (Zhang et al., 2025). Through the authoring policy, we guide the agent to generalize demonstration- specific inputs into task-dependent descriptions and describe UI targets semantically. The policy instructs the agent to keep phases together when they share state, task-derived data, quantities, or required ordering, and to retain selection rules, calculations, dependencies, requested scope, stop conditions, visible completion checks, and supported recovery. Skills must preserve visible GUI interaction and include all instructions needed for execution without the source recording. Before completion, the agent calls finalize extraction to validate output structure and required tool use. Checks cover task-file access, required skill structure, a recorded write followed by a read of each final skill, and metadata inspection before video sampling. Successful finalization records skill hashes, inspected source paths, validation results, and run provenance. These checks verify output structure and required tool use; they do not verify skill correctness or task success. For downstream evaluation, the executor receives the target task instruction and generated skill. It selects GUI actions using this guidance together with its observations and action history, without accessing the source recording. Sections 4.4 and 4.3 evaluate reuse across execution models and task instances. Appendix D.4 details the authoring policy and output requirements. 4EXPERIMENTS We organize our experiments around the following four research questions. RQ1 tests whether the induced skills improve execution on the demonstrated tasks. RQ2 and RQ3 then examine two dimensions of reuse: changes to task requirements and changes to the execution model, respectively. RQ4 investigates how the available evidence and induction prompt affect downstream utility. A case study illustrates how induced skills handle dependent subgoals in a multistep workflow. 4.1EXPERIMENTAL SETUP Benchmarks. We evaluate on the 64 MyPCBench Bounded Action tasks and 286 OSWorld- Verified tasks in WorkDemo-350 using official task evaluators. On MyPCBench, the execution agent has access to both GUI actions and Bash commands. Benchmark selection and demonstration screening are described in Section 3.2. OSWorld results apply to this curated subset. Appendix E.3 describes our OSWorld execution harness and evaluator configuration. Skill-based Evaluation. We evaluate skills as plug-in guidance by appending their text as a clearly delimited block to the end of the executorâs base system prompt. All text-guidance methods use the same <task skills> wrapper; Appendix D.5 gives the template and injected contents. Separating procedural guidance from the execution loop supports reuse across execution models and integra- tion with different agent harnesses. We therefore use shared execution settings for each model and benchmark, varying only the appended guidance to control for differences in native agent pipelines. We compare Work2Skill with No Skill, action semantics from ShowUI-Aloha (Aloha) (Zhang et al., 2026e), and tutorial text from Demo2Tutorial (D2T) (Bai et al., 2026). The demonstration-based methods use the same source episodes, while their input modalities and evidence representations follow each methodâs design. RQ1 thus compares their overall ability to produce useful guidance un- der shared execution conditions; RQ4 separately examines evidence and induction-prompt choices within Work2Skill while holding the model and skill-authoring policy fixed. Models and Metrics. We use GPT-5.5 (OpenAI, 2026a), Opus-4.8 (Anthropic, 2026a), Qwen3.5- 35B-A3B, and Qwen3.5-9B (Qwen Team, 2026). By default, the same model induces skills and executes tasks with them; RQ3 varies this pairing. We report mean evaluation score (Score, 0â 100), full-credit success rate (SR), mean execution steps, and mean skill-induction token usage. Appendix E details task selection, metrics, baseline adaptations, and execution settings. 4.2RQ1: DOES WORK2SKILL IMPROVE DOWNSTREAM GUI EXECUTION? To evaluate the effectiveness of induced skills in downstream GUI execution, we compare Work2Skill with No Skill and the text-adapted ShowUI-Aloha and D2T baselines on the demon- strated tasks under shared execution settings (Table 2). Relative to the higher-scoring of the two 6 Preprint Table 2: Performance and efficiency on MyPCBench and the evaluated OSWorld-Verified subset. Bold indicates the best Score or SR within each model and benchmark, and blue shading denotes our method. Score is the mean evaluation score (0â100), and SR is the percentage of tasks receiving full credit on either benchmark. Steps denotes average execution steps, and Tok. (k) denotes average token usage for skill induction, in thousands. Aloha denotes ShowUI-Aloha. Method GPT-5.5Opus-4.8Qwen3.5-35B-A3BQwen3.5-9B Score SR (%) Steps Tok. (k) Score SR (%) Steps Tok. (k) Score SR (%) Steps Tok. (k) Score SR (%) Steps Tok. (k) MyPCBench (N = 64) No Skill 74.73 23.44 40.89â81.53 42.19 40.58â38.389.3856.28â24.276.2570.92â Aloha77.48 37.50 41.33 170.65 79.27 34.38 37.94 256.23 37.69 12.50 67.47 243.86 26.613.1379.77 235.36 D2T75.25 37.50 46.13 177.32 81.28 42.19 36.84 234.73 40.619.3863.45 211.98 27.756.2570.84 180.48 Ours85.2748.4447.38334.5286.2050.0036.03300.2144.5620.3146.69245.6334.1112.5069.30249.87 OSWorld-Verified (N = 286) No Skill 74.47 70.98 19.05â79.23 77.97 26.37â32.52 32.17 66.61â27.48 26.92 70.02â Aloha72.57 69.93 21.61 146.98 78.91 76.57 18.04 132.24 31.08 30.42 58.01 369.27 24.81 24.13 72.51 319.98 D2T80.32 76.22 21.46 134.93 80.66 78.67 17.65 165.80 33.93 32.52 65.52 219.82 26.45 25.52 73.84 175.27 Ours81.9778.6725.78185.1982.3479.3720.15238.6636.3636.0165.89218.8833.4132.5268.13261.08 demonstration-based baselines for each execution model, Work2Skill improves Score by 3.95â7.79 points on MyPCBench and 1.65â6.96 points on the evaluated OSWorld-Verified subset. For GPT-5.5 on MyPCBench, SR reaches 48.44%, compared with 37.50% for both Aloha and D2T. Compared with No Skill, Work2Skill improves the average score by 4.67â10.54 points and the success rate by 6.25â25.00 percentage points on MyPCBench. On the evaluated OSWorld-Verified subset, the corresponding gains are 3.11â7.50 score points and 1.40â7.69 percentage points. Efficiency gains depend on the model: Work2Skill uses fewer execution steps than No Skill for Opus-4.8 and both Qwen models, but more for GPT-5.5. Skill-induction token usage is also not consistently lower than that of the demonstration-based baselines. Thus, Work2Skill achieves the highest Score and SR among the compared methods across both task sets and all four executors, with model-dependent execution and induction costs. 4.3RQ2: DO SKILLS GENERALIZE TO NEW TASK INSTANCES? To assess whether induced skills generalize beyond the demonstrated instance, we evaluate the origi- nal skills on tasks with changed requirements within the same application, without further induction. MyPCBench Jitter contains one paired variant of each of the 64 original tasks, changing entities, pa- rameters, or constraints within the same application context. For example, a flight-booking variant changes the destination, travel date, and seat preference. GPT-5.6-sol assists in revising task in- structions and evaluation criteria, followed by human verification (Appendix E.1). Each of the four models directly reuses the skills induced for the original tasks on their paired variants. Table 3: Performance on 64 MyPCBench Jitter tasks. Bold indicates the best Score or SR within each execution model, and blue shading denotes our method. Smaller annotations report changes relative to No Skill, computed from the displayed two-decimal values. For Score and SR, upward arrows indicate improvements; for Steps, downward arrows indicate fewer execution steps. Method GPT-5.5Opus-4.8Qwen3.5-35B-A3BQwen3.5-9B ScoreSR (%)StepsScoreSR (%)StepsScoreSR (%)StepsScoreSR (%)Steps No Skill 73.4537.5052.4183.3045.3151.9742.9412.5067.6627.806.2582.33 Aloha75.50 â2.05 42.19 â4.69 51.45 â0.96 82.66 â0.64 43.75 â1.56 48.17 â3.80 41.19 â1.75 12.50 â0.00 63.14 â4.52 31.97 â4.17 10.94 â4.69 78.63 â3.70 D2T73.03 â0.42 35.94 â1.56 51.94 â0.47 81.44 â1.86 45.31 â0.00 48.48 â3.49 46.92 â3.98 14.06 â1.56 66.31 â1.35 30.66 â2.86 12.50 â6.25 81.66 â0.67 Ours81.75 â8.30 46.88 â9.38 49.78 â2.63 87.23 â3.93 54.69 â9.38 46.53 â5.44 51.58 â8.64 21.88 â9.38 65.17 â2.49 36.25 â8.45 12.50 â6.25 79.53 â2.80 In Table 3, Work2Skill improves the average score over No Skill by 3.93â8.64 points and the success rate by 6.25â9.38 percentage points. Compared with the stronger demonstration-based baseline for each executor, the Score advantage is 4.28â6.25 points. Work2Skill also achieves the highest success rate for three executors, while tying D2T for Qwen3.5-9B. Average execution steps also decrease by 2.49â5.44 relative to No Skill across the four execution models. These gains across all four 7 Preprint executors support reuse beyond the demonstrated instance: the induced skills remain useful when entities, parameters, or constraints change within the tested applications. 4.4RQ3: DO SKILLS TRANSFER ACROSS MODELS? Execution Induction GPT-5.5 Qwen3.5-9B Score GPT-5.5 85.27 Î +10.54 75.81 Î +1.08 Qwen3.5-9B 36.66 Î +12.39 34.11 Î +9.84 Success Rate (%) GPT-5.5 48.44 Î +25.00 39.06 Î +15.62 Qwen3.5-9B 14.06 Î +7.81 12.50 Î +6.25 Figure 4: Cross-model skill transfer on 64 MyPCBench tasks. Rows denote induction mod- els; columns denote executors. â denotes the gain over the same executorâs No Skill baseline, in Score points or SR percentage points, calcu- lated from displayed two-decimal values. To evaluate whether induced skills transfer across models, we separate induction from execution and test guidance produced by one model with a different executor. We cross GPT-5.5 and Qwen3.5-9B as in- duction and execution models, freezing each induced skill before downstream evaluation (Figure 4). Skills induced by GPT-5.5 outperform those induced by Qwen3.5-9B under both executors.The ad- vantage is larger with GPT-5.5 execution, reach- ing 9.46 score points and 9.38 percentage points in success rate, but remains positive after transfer to Qwen3.5-9B. All four pairings improve on the No Skill baseline for the same executor. Perfor- mance remains substantially lower with Qwen3.5-9B execution regardless of the induction model. Thus, the skills transfer usefully between these two models, but stronger induction does not close the performance gap between executors on these tasks. ScoreSuccess rate (%)Steps 020406080 Full Evidence 85.27 48.44 47.38 w/o Indexed Images 74.64 32.81 46.23 w/o UI Annotations 80.25 34.38 42.58 w/o KeyLogs 78.03 26.56 41.06 Video-only 75.81 28.13 41.05 UI-Grounded Action Text Only 79.09 37.50 34.56 (a) Ablation of evidence used for skill induction. Score 0 50 100 81.30 Pipeline 85.27 Ours Success rate (%) 0 30 60 42.19 Pipeline 48.44 Ours Steps 0 30 60 48.98 Pipeline 47.38 Ours Skill-induction tokens (10 4 ) 0 20 40 29.68 Pipeline 33.45 Ours (b) Comparison with the pipeline prompt. Figure 5: Ablation and pipeline-prompt comparison on 64 MyPCBench tasks using GPT-5.5 with high reasoning effort. In (a), Full Evidence uses all available modalities; UI-Grounded Action Text Only combines recorded actions with target UI descriptions. w/o KeyLogs also removes overlays from indexed images. Score, SR, and Steps measure downstream execution; Tokens reports mean skill-induction token usage per task in units of 10 4 . 4.5RQ4: HOW DO EVIDENCE AND INDUCTION PROMPTS AFFECT SKILL INDUCTION? To assess how evidence and prompts affect downstream skill utility, we ablate the available input modalities and separately compare the Work2Skill and Pipeline Prompt system prompts. For the evidence ablation, we vary induction evidence on 64 MyPCBench tasks with GPT-5.5, keeping reasoning effort (high) and the iteration limit fixed. UI annotations contain metadata and descriptions derived from the corresponding action targets in the AX tree. The w/o KeyLogs condition removes both raw logs and added image overlays while retaining the action trace and its UI target information. Appendix E.6 details the inputs and prompt adaptations. Full Evidence scores 85.27 and fully solves 31 tasks (Figure 5a). Removing indexed images, UI annotations, or KeyLogs reduces Score by 5.02â10.63 points and SR by 14.06â21.88 percentage points. The largest drops are in Score without indexed images and SR without KeyLogs. UI-Grounded Action Text Only uses ordered action descriptions with target UI semantics. It outperforms Video-only and uses the fewest steps; both trail Full Evidence in Score and SR. Fewer steps do not necessarily indicate better task completion; Score and SR remain essential for interpreting execution efficiency. 8 Preprint With the evidence held fixed, we next test whether a prescribed sequence of induction stages helps the agent use it. Our concise default prompt leaves detailed evidence exploration and skill revision to the agent. Pipeline Prompt specifies the induction stages and their order: read the task and artifacts, inspect images and video, then write, review, and finalize skills. Only the system prompt changes; the model, evidence, tools, authoring policy, and output format remain fixed (Appendix D.3). Both prompts allow further evidence inspection. In Figure 5b, Work2Skill improves Score by 3.97 points and SR by 6.25 percentage points, while reducing average execution steps from 48.98 to 47.38, at 12.7% higher skill-induction token usage. In this setting, Full Evidence yields the best Score and SR among the tested input configurations, while the Work2Skill prompt outperforms the prescribed pipeline at higher induction cost. Together, these comparisons show how evidence and prompt design affect the downstream utility of generated skills. Configure main reservation Open and inspect three alternative restaurants (incomplete evidence) Not completed Confirm the reservation Verify final state Configure main reservation Open and inspect three alternative restaurants (incomplete evidence) Not completed Confirm the reservation Verify final state Configure main reservation Open and inspect three alternative restaurants Save the top two (favorites) Add note to Cotogna and confirm Confirm the reservation Verify final state Evaluation Rubric (R1 âR8) R1ďźDate + table for 10 R2ďźCotogna booked + note R3ďźOpen 3 alternatives R4ďźView each detail page R5ďź Save top 2 R6ďźConfirmation number R8ďźCorrect final details ShowUI-Aloha (action-only) â Ă Ă Demo2Tutorial(text-only) Ă Work2Skill (Ours) Reserve Cotogna for 10 next Friday, inspect three alternative Scranton restaurants, save the best two, add note âDundies after-partyâ, confirm the reservation, and verify results. Task Instruction: ShowUI- Aloha Demo2T utorial Work2 Skill â â â Score: 55/100 Score: 100/100 Score: 55/100 Missing step Missing step R7ďźRefresh persistence check â â â Ă Ă â Ă Ă â Ă Ă â â â â â â â Figure 6: Execution trajectories (top) and rubric outcomes (bottom) for a TableFind reservation with fallbacks, using GPT-5.5 with each methodâs extracted guidance. Both baselines leave fallback handling and persistence checks incomplete (55/100); Work2Skill satisfies all eight requirements (100/100). Appendix F.2 provides the detailed rubric and execution analysis. 4.6CASE STUDY To illustrate how extracted skill guidance affects execution, we compare GPT-5.5 supplied with out- puts from ShowUI-Aloha, D2T, or Work2Skill through the shared text interface (Section 4.1). Each method extracts its guidance from the same human demonstration. Figure 6 shows a Cotogna book- ing for ten people with three fallback restaurants and two favorites. The fallbacks must be handled before confirmation, while preserving the date, party size, and requested note. With ShowUI-Aloha or D2T guidance, the executor completes the primary booking but leaves fallback inspection, favorit- ing, and persistence checks incomplete, scoring 55/100 in both conditions. With the skill induced by Work2Skill, it satisfies all eight rubric items (100/100). The generated skill makes intermedi- ate requirements explicit: it rechecks fields that may reset during navigation, excludes the primary restaurant from the fallback set, and defers final confirmation until the fallback steps are complete. The skill preserves dependent subgoals across navigation and checks both the reservation and saved favorites before declaring completion (Appendix F.2). 5CONCLUSION We presented Work2Skill, a framework for capturing human GUI experience and converting it into reusable agent skills. WorkRecorder collects aligned multimodal evidence, WorkDemo-350 orga- nizes 350 curated demonstrations, and agentic skill induction combines on-demand evidence inspec- tion with iterative skill writing and revision. Across four execution models, Work2Skill achieves the 9 Preprint highest scores and success rates among the compared methods on MyPCBench Bounded Action and the evaluated OSWorld-Verified subset. Further experiments support reuse under changed task requirements within the same application and transfer between the tested models. These results highlight the value of aligned human demonstrations and flexible evidence exploration for building reusable procedural guidance that computer-use agents can apply through a shared text interface without accessing the original recordings or other demonstration evidence. REFERENCES Reyna Abhyankar, Qi Qi, and Yiying Zhang. Osworld-human: Benchmarking the efficiency of computer-use agents. Proceedings of Machine Learning and Systems, 8:482â494, 2026. Saaket Agashe, Kyle Wong, Vincent Tu, Jiachen Yang, Ang Li, and Xin Eric Wang.Agent S2: A Compositional Generalist-Specialist Framework for Computer Use Agents.CoRR, abs/2504.00906, 2025. doi: 10.48550/ARXIV.2504.00906. Anthropic. Introducing Claude Opus 4.8, May 2026a. URL https://w.anthropic.com/ news/claude-opus-4-8. May 28, 2026. Accessed September 22, 2026. Anthropic. How to create custom skills, 2026b. URL https://support.claude.com/en/ articles/12512198-how-to-create-custom-skills. Online documentation, âRecord a skillâ section. Accessed September 20, 2026. Zechen Bai, Zhiheng Chen, Yiqi Lin, Kevin Qinghong Lin, Difei Gao, Xiangwu Guo, Xin Wang, and Mike Zheng Shou.Demo2Tutorial: From Human Experience to Multimodal Software Tutorials. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2026, Denver, CO, USA, June 3-7, 2026, p. 29588â29597, 2026.URL https: //openaccess.thecvf.com/content/CVPR2026/html/Bai_Demo2Tutorial_From_ Human_Experience_to_Multimodal_Software_Tutorials_CVPR_2026_paper.html. Aarash Feizi, Shravan Nayak, Xiangru Jian, Kevin Qinghong Lin, Kaixin Li, Rabiul Awal, Xing Han L ` u, Johan S Obando Ceron, Juan A Rodriguez, Nicolas Chapados, et al. Grounding computer use agents on human demonstrations. In International Conference on Learning Representations, volume 2026, p. 5181â5215, 2026. Lawrence Keunho Jang, Andrew Keunwoo Jang, Jing Yu Koh, and Ruslan Salakhutdinov. MyPCBench:A Benchmark for Personally Intelligent Computer-Use Agents.CoRR, abs/2606.16748, 2026. doi: 10.48550/ARXIV.2606.16748. Keliang Li, Yansong Li, Hongze Shen, Mengdi Liu, Hong Chang, and Shiguang Shan. LensWalk: Agentic Video Understanding by Planning How You See in Videos. CoRR, abs/2603.24558, 2026. doi: 10.48550/ARXIV.2603.24558. Yinheng Li, Hailey Hultquist, Justin Wagle, and Kazuhito Koishida. Instruction Agent: Enhancing Agent with Expert Demonstration. CoRR, abs/2509.07098, 2025. doi: 10.48550/ARXIV.2509. 07098. Guangyi Liu, Pengxiang Zhao, Liang Liu, Zhiming Chen, Yuxiang Chai, Yaozhen Liang, Wen- Hao Wang, Siheng Chen, Zhengxi Lu, Shuai Ren, Hao Wang, Shibo He, Yong Liu, and Wen- chao Meng. LearnAct: Few-shot mobile GUI agent with a unified demonstration benchmark. In Maria Liakata, Viviane P. Moreira, Jiajun Zhang, and David Jurgens (eds.), Findings of the Association for Computational Linguistics: ACL 2026, p. 29820â29843, San Diego, Califor- nia, United States, July 2026. Association for Computational Linguistics. ISBN 979-8-89176- 395-1. doi: 10.18653/v1/2026.findings-acl.1491. URL https://aclanthology.org/2026. findings-acl.1491/. Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. DrVideo: Document Retrieval Based Long Video Understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2025, Nashville, TN, USA, June 11-15, 2025, p. 18936â18946, 2025. doi: 10.1109/CVPR52734.2025.01764. 10 Preprint Qirui Mi, Zhijian Ma, Mengyue Yang, Haoxuan Li, Yisen Wang, Haifeng Zhang, and Jun Wang. ProcMEM: Learning Reusable Procedural Memory from Experience via Non-Parametric PPO for LLM Agents. CoRR, abs/2602.01869, 2026. doi: 10.48550/ARXIV.2602.01869. Shravan Nayak, Xiangru Jian, Kevin Qinghong Lin, Juan A. Rodr Ě Äąguez, Montek Kalsi, Nico- las Chapados, M. Tamer Ě Ozsu, Aishwarya Agrawal, David V Ě azquez, Christopher Pal, Perouz Taslakian, Spandana Gella, and Sai Rajeswar.UI-Vision: A Desktop-centric GUI Bench- mark for Visual Perception and Interaction. In Forty-second International Conference on Ma- chine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025, 2025. URL https: //proceedings.mlr.press/v267/nayak25a.html. Nex-AGI. NexAU, 2026. URL https://github.com/nex-agi/NexAU. GitHub repository. Ac- cessed September 20, 2026. Dang Nguyen, Jian Chen, Yu Wang, Gang Wu, Namyong Park, Zhengmian Hu, Hanjia Lyu, Junda Wu, Ryan Aponte, Yu Xia, Xintong Li, Jing Shi, Hongjie Chen, Viet Dac Lai, Zhouhang Xie, Sungchul Kim, Ruiyi Zhang, Tong Yu, Mehrab Tanjim, Nesreen K. Ahmed, Puneet Mathur, Se- unghyun Yoon, Lina Yao, Branislav Kveton, Jihyung Kil, Thien Huu Nguyen, Trung Bui, Tianyi Zhou, Ryan A. Rossi, and Franck Dernoncourt. GUI agents: A survey. In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar (eds.), Findings of the Associa- tion for Computational Linguistics: ACL 2025, p. 22522â22538, Vienna, Austria, July 2025. Association for Computational Linguistics. ISBN 979-8-89176-256-5. doi: 10.18653/v1/2025. findings-acl.1158. URL https://aclanthology.org/2025.findings-acl.1158/. Jingwei Ni, Yihao Liu, Xinpeng Liu, Yutao Sun, Mengyu Zhou, Pengyu Cheng, Dexin Wang, Er- chao Zhao, Xiaoxi Jiang, and Guanjun Jiang. Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills. CoRR, abs/2603.25158, 2026. doi: 10.48550/ARXIV.2603.25158. OpenAI.Introducing GPT-5.5, April 2026a.URL https://openai.com/index/ introducing-gpt-5-5/. April 23, 2026. Accessed September 22, 2026. OpenAI.Record & Replay, 2026b.URL https://learn.chatgpt.com/docs/extend/ record-and-replay. Online documentation. Accessed September 20, 2026. Zhiyuan Peng, Xin Yin, Chenhao Ying, Zhe Cui, Zixiang Ding, Zhenhua Liu, Jiang Wu, and Yuan Luo. EvoClawBench: Can Agents Learn Reusable Skills from Their Own Runs? CoRR, abs/2607.09711, 2026. doi: 10.48550/ARXIV.2607.09711. Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, Wanjun Zhong, Kuanye Li, Jiale Yang, Yu Miao, Woyu Lin, Longx- iang Liu, Xu Jiang, Qianli Ma, Jingyu Li, Xiaojun Xiao, Kai Cai, Chuang Li, Yaowei Zheng, Chaolin Jin, Chen Li, Xiao Zhou, Minchao Wang, Haoli Chen, Zhaojian Li, Haihua Yang, Haifeng Liu, Feng Lin, Tao Peng, Xin Liu, and Guang Shi. UI-TARS: Pioneering Automated GUI Inter- action with Native Agents. CoRR, abs/2501.12326, 2025. doi: 10.48550/ARXIV.2501.12326. Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URL https://qwen. ai/blog?id=qwen3.5. Mirac Suzgun, Mert Y Ě uksekg Ě on Ě ul, Federico Bianchi, Dan Jurafsky, and James Zou. Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory. In Proceedings of the 19th Confer- ence of the European Chapter of the Association for Computational Linguistics, EACL 2026 - Volume 1: Long Papers, Rabat, Morocco, March 24-29, 2026, p. 7080â7106, 2026. doi: 10.18653/V1/2026.EACL-LONG.333. Fei Tang, Haolei Xu, Hang Zhang, Siqi Chen, Xingyu Wu, Yongliang Shen, Wenqi Zhang, Guiyang Hou, Zeqi Tan, Yuchen Yan, Kaitao Song, Jian Shao, Weiming Lu, Jun Xiao, and Yueting Zhuang. A Survey on (M)LLM-Based GUI Agents. CoRR, abs/2504.13865, 2025. doi: 10.48550/ARXIV. 2504.13865. Xinyuan Wang, Bowen Wang, Dunjie Lu, Junlin Yang, Tianbao Xie, Junli Wang, Jiaqi Deng, Xi- aole Guo, Yiheng Xu, Chen Henry Wu, Zhennan Shen, Zhuokai Li, Ryan Li, Xiaochuan Li, Junda Chen, Boyuan Zheng, Peihang Li, Fangyu Lei, Ruisheng Cao, Yeqiao Fu, Dongchan 11 Preprint Shin, Martin Shin, Jiarui Hu, Yuyan Wang, Jixuan Chen, Yuxiao Ye, Danyang Zhang, Yipu Wang, Heng Wang, Diyi Yang, Victor Zhong, Y. Charles, Zhilin Yang, and Tao Yu. Open- CUA: Open Foundations for Computer-Use Agents. In Advances in Neural Information Process- ing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, November 30 - De- cember 5, 2025, 2025a. URL http://papers.nips.c/paper_files/paper/2025/hash/ c7ae529e945226b0d52ea4ac478c4f3-Abstract-Conference.html. Ziyang Wang, Honglu Zhou, Shijie Wang, Junnan Li, Caiming Xiong, Silvio Savarese, Mohit Bansal, Michael S Ryoo, and Juan Carlos Niebles. Active video perception: Iterative evidence seeking for agentic long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 9088â9099, 2026. Zora Zhiruo Wang, Apurva Gandhi, Graham Neubig, and Daniel Fried. Inducing Programmatic Skills for Agentic Tasks. CoRR, abs/2504.06821, 2025b. doi: 10.48550/ARXIV.2504.06821. Peng Xia, Jianwen Chen, Hanyang Wang, Jiaqi Liu, Kaide Zeng, Yu Wang, Siwei Han, Yiyang Zhou, Xujiang Zhao, Haifeng Chen, Zeyu Zheng, Cihang Xie, and Huaxiu Yao. SkillRL: Evolv- ing Agents via Recursive Skill-Augmented Reinforcement Learning. CoRR, abs/2602.08234, 2026. doi: 10.48550/ARXIV.2602.08234. Yixiong Xiao, Lang An, Hucheng Yang, Pinxue Ma, Yongquan Chen, Jingjia Cao, Yusai Zhao, Ting Wang, Ting Liu, Siqi Bao, et al. Omegause-sop: Sop engineering for professional computer use from human demonstrations. arXiv preprint arXiv:2609.02149, 2026. Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu.OSWorld: Bench- marking Multimodal Agents for Open-Ended Tasks in Real Computer Environments.In Advances in Neural Information Processing Systems 37: Annual Conference on Neural In- formation Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024, 2024.URL http://papers.nips.c/paper_files/paper/2024/hash/ 5d413e48f84dc61244b6be550f1cd8f5-Abstract-Datasets_and_Benchmarks_Track. html. Honjar Xing, Jefferson Lin, and Henry Lieberman. How Should Agents Read Demonstrations? Hierarchical Structure Beats Flat Action Logs. CoRR, abs/2606.20978, 2026. doi: 10.48550/ ARXIV.2606.20978. Kaisen Yang, Zheng Jiang, Yuzhao Peng, Houde Qian, Boshi Zhang, Youjie Zheng, Shijin Hong, Qingle Liu, Ruoyu Han, Bohan Lyu, Bingxiang He, Eren Cai, Calvin Xiao, and Qinhuai Na. Scalable Behaviour Cloning on Browser Using via Skill Distillation. CoRR, abs/2606.32014, 2026a. doi: 10.48550/ARXIV.2606.32014. Shidong Yang, Ziyu Ma, Tongwen Huang, Xucong Wang, Renda Li, Yiming Hu, Yong Wang, and Xiangxiang Chu. SkillForge: Evolving Verifiable Skills for Reinforcement Learning Agents. CoRR, abs/2608.24747, 2026b. doi: 10.48550/ARXIV.2608.24747. Yutao Yang, Junsong Li, Qianjun Pan, Bihao Zhan, Yuxuan Cai, Lin Du, Jie Zhou, Kai Chen, Qin Chen, Xin Li, Bo Zhang, and Liang He. AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution. CoRR, abs/2603.01145, 2026c. doi: 10.48550/ARXIV.2603.01145. Zeyuan Yang, Delin Chen, Xueyang Yu, Maohao Shen, and Chuang Gan. VCA: Video Curious Agent for Long Video Understanding. In IEEE/CVF International Conference on Computer Vision, ICCV 2025, Honolulu, HI, USA, October 19-25, 2025, p. 20168â20179, 2025. doi: 10.1109/ICCV51701.2025.01876. Yiwen Yin, Yu Mei, Chun Yu, Toby Jia-Jun Li, Aamir Khan Jadoon, Sixiang Cheng, Weinan Shi, Mohan Chen, and Yuanchun Shi. From Operation to Cognition: Automatic Modeling Cognitive Dependencies from User Demonstrations for GUI Task Automation. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, CHI 2025, YokohamaJapan, 26 April 2025- 1 May 2025, p. 684:1â684:24, 2025. doi: 10.1145/3706598.3713356. 12 Preprint Huaying Yuan, Zheng Liu, Junjie Zhou, Hongjin Qian, Ji-Rong Wen, and Zhicheng Dou. VideoDeepResearch:Long Video Understanding With Agentic Tool Using.CoRR, abs/2506.10821, 2025. doi: 10.48550/ARXIV.2506.10821. Barry Zhang,Keith Lazuka,and Mahesh Murag.Equipping agents for the real world with Agent Skills, 2025.URL https://w.anthropic.com/engineering/ equipping-agents-for-the-real-world-with-agent-skills. Anthropic Engineering, October 16, 2025. Accessed September 21, 2026. Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue, and Wenya Wang. MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents. CoRR, abs/2602.02474, 2026a. doi: 10.48550/ARXIV.2602.02474. Kangning Zhang, Shuai Shao, Qingyao Li, Jianghao Lin, Lingyue Fu, Shijian Wang, Wenxiang Jiao, Yuan Lu, Weiwen Liu, Weinan Zhang, and Yong Yu. MMSkills: Towards Multimodal Skills for General Visual Agents. CoRR, abs/2605.13527, 2026b. doi: 10.48550/ARXIV.2605.13527. Qizheng Zhang, Changran Hu, Shubhangi Upasani, Boyuan Ma, Fenglu Hong, Vamsidhar Kama- nuru, Jay Rainton, Chen Wu, Mengmeng Ji, Hanchen Li, et al. Agentic context engineering: Evolving contexts for self-improving language models. In International Conference on Learning Representations, volume 2026, p. 86069â86100, 2026c. Weijia Zhang, Kunlun Zhu, Zeyi Liu, Yinting Chen, Tianyi Ma, Jiateng Liu, Jiaxun Zhang, Bingxuan Li, Xiangru Tang, Heng Ji, and Jiaxuan You. CUADebug: Diagnosing and Repairing Computer- Use Agent Failures. CoRR, abs/2608.02643, 2026d. doi: 10.48550/ARXIV.2608.02643. Yichun Zhang, Xiangwu Guo, Yauhong Goh, Jessica Hu, Zhiheng Chen, Xin Wang, Difei Gao, and Mike Zheng Shou. ShowUI-Aloha: Human-Taught GUI Agent. CoRR, abs/2601.07181, 2026e. doi: 10.48550/ARXIV.2601.07181. Zirui Zhao, Jun Hao Liew, Yan Yang, Wenzhuo Yang, Ziyang Luo, Doyen Sahoo, Silvio Savarese, and Junnan Li.GPA: Learning GUI Process Automation from Demonstrations.CoRR, abs/2604.01676, 2026. doi: 10.48550/ARXIV.2604.01676. Boyuan Zheng, Michael Y. Fatemi, Xiaolong Jin, Zora Zhiruo Wang, Apurva Gandhi, Yueqi Song, Yu Gu, Jayanth Srinivasa, Gaowen Liu, Graham Neubig, and Yu Su. SkillWeaver: Web Agents can Self-Improve by Discovering and Honing Skills. CoRR, abs/2504.07079, 2025. doi: 10. 48550/ARXIV.2504.07079. 13 Preprint ARECORDER IMPLEMENTATION A.1ARCHITECTURE AND DEPLOYMENT WorkRecorder consists of a native macOS recording application, a local service for capture coor- dination and storage, and accessibility providers. In local macOS mode, the application records the selected screen region and input activity, with available accessibility evidence supplied by the macOS provider. For a remote Linux workflow, recording remains on the Mac, while an AT-SPI collector runs inside the target desktop session. The Linux launcher connects this collector to the local service through an SSH tunnel; the remote-desktop client supplies the displayed workflow. Task initialization and official benchmark evaluation are handled separately from recording. Video and input capture. The macOS application records a selected screen region through ScreenCaptureKit and encodes video with H.264. CoreGraphics event taps supply mouse and key- board activity, which is also grouped into semantic operations. The operator can pause, resume, stop, and review a recording. Full keystroke capture is disabled by default; when enabled, it records key presses and releases, characters, modifiers, and timestamps. Otherwise, ordinary typing is repre- sented through aggregate typing events. The capture policy therefore determines the detail available in the input stream. Accessibility evidence. The Linux collector supports X11/Xorg desktops and uses AT-SPI to ob- tain Accessibility Trees, interface events, and target UI elements. Its default periodic tree sampler runs every two seconds, with limits of six levels and 400 nodes; these are implementation defaults rather than dataset collection settings. Events may reference the latest completed tree snapshot, with its sampling timestamps. Target-element queries return available names, roles, and bounds at a screen location. The recorder maps positions in the displayed remote-desktop region to Linux desktop coordinates for these queries. Availability depends on application accessibility support and sampling limits. Target-element association. When resolving clicks from accessibility snapshots, the resolver considers nodes whose bounds contain the click coordinate. It ranks these candidates by node confi- dence and interactive role, with an area penalty favoring smaller controls. Resolution records retain the selected source, confidence, and evidence anchor; uncertain cases can remain flagged for review. The remote Linux probing path is triggered by clicks and drags. Configured remote-display geome- try maps local pointer positions to Linux desktop pixels; positions outside the displayed desktop do not trigger a point probe. The AT-SPI watcher combines bounded point queries with nearby focus, selection, and state-change events. Its point-query ranking prioritizes active windows, named inter- active controls, other named nodes, greater depth, and smaller area, in that order. Event correlation normally checks whether the target bounds contain the pointer position, with additional handling for transient controls such as menus that disappear after selection. Pointer-resolution records retain the originating pointer sequence identifier and distinguish resolved, coarse, unresolved, and error or timeout outcomes. During export, recorded probe results are associated with actions using their triggerâs monotonic timestamp and action kind, with a nearest-action fallback within 2.5 seconds. In normal export mode, a missing click or drag target can also be recovered from the nearest specific accessibility event within a 3-second window. This backfill excludes generic page or window focus and uses each candidate event at most once. For drags, the matching anchor is the drag start, computed from the release time and recorded duration. Backfilled records retain their source, event identifier, and time offset. These are heuristic temporal associations; the matching windows are implementation tolerances, not measured accuracy bounds. Keyboard input does not trigger this Linux target probe. AX snapshots are attached independently as interface context. A.2KEYFRAME SELECTION Capture triggers. WorkRecorder selects keyframes primarily around recorded operations and in- terface transitions. Recording start, stop, pause, and resume provide boundary frames, while users can explicitly mark additional frames. Routine candidates are triggered by clicks, drag releases, 14 Preprint scroll completion, recognized hotkeys, command submission, and window or tab switches. A typ- ing pause can trigger a frame when text is submitted or a typing milestone is reached; the default milestone requires at least eight characters or input events in the typing segment. Individual key presses and releases do not independently trigger screenshots. Additional enabled policies can re- quest frames at scroll milestones or when a provider reports an error or warning surface. Action context and capture timing. For clicks, hotkeys, and window or tab switches, the recorder can reuse a recent preceding keyframe as before-action context, with a default freshness limit of 1.5 seconds. A new before-action capture requires an explicit pre-action event; otherwise this context can remain unavailable. Result frames use a bounded settling window to allow the interface to respond. With adaptive polling enabled, capture can proceed when a provider reports readiness or when visual or text observations stabilize after a detected change; the default stability rule requires two consecutive stable samples. Waiting limits depend on the action type. If settling times out, the policy can capture the available state and flag it for review, or suppress the candidate. Thus, a result-frame link does not by itself establish that the interface has finished updating. Candidate filtering. The planner removes duplicate triggers within an actionâs timing plan. The scheduler applies a minimum gap between separate plans, limits result frames per action, and en- forces capture budgets. When change or quality signals are available, it can also reject candidates with insufficient change or information gain. Defaults include a 500-ms gap, at most two result- related candidates per action, and a session budget of 64 keyframes; required boundary frames and explicit user marks have exemptions. These are implementation defaults, not verified settings for every collected demonstration. Retained keyframes carry identifiers and timestamps, while trig- ger decisions retain acceptance or rejection reasons. Frame reuse and candidate filtering mean that actions and keyframes need not have a one-to-one correspondence. Exported image counts are dis- tinguished from recorded keyframe counts in Appendix B. A.3TIME ALIGNMENT The recorder maps evidence onto the stitched videoâs timeline. Native input uses a monotonic times- tamp anchored to the first captured frame of each recording segment, rather than the time at which recording was requested. Remote Linux events use wall-clock timestamps, which are projected onto the same timeline using the corresponding first-frame wall-clock anchors and segment durations. Segment offsets account for pause/resume boundaries; remote events falling within a pause have no corresponding video frame. The Linux launcher checks host clock agreement before recording and rejects a difference greater than two seconds. This check is a deployment condition, not a measured alignment-error bound. Legacy recordings without first-frame wall-clock anchors use the session start as a fallback. BEVIDENCE SCHEMA B.1EPISODE FILES FOR SKILL INDUCTION The induction agent reads a per-task episode directory containing the task and its recorded ev- idence.The benchmark converter creates three core files: task.json, episode.json, and trace/actions.jsonl. Input logs, video, summaries, and images are included when available (Table 4). The task file must contain the authoritative instruction and task inputs; a placeholder with only an inferred task ID is insufficient for induction. The converter excludes the grading field from this file. The episode.json manifest identifies the task and recording, records duration and evidence counts, and indexes the available files. Evidence paths are relative to the episode directory. Its source entry records the original export location for provenance; that location is not an additional input root for the agent. 15 Preprint Table 4: Files exposed to the induction agent in a converted episode. The first three files form the core structure; the remaining artifacts are included when available. File or directoryContent task.jsonTask instruction and associated task metadata. episode.jsonFile references, recording metadata, and counts. trace/actions.jsonlNormalized action records in temporal order. trace/keylog.jsonlRecorded input-event log. video.mp4Recording video. summary.mdTextual operation summary. evidence/Evidence images, with available image manifests and action- association indices. B.2ACTION RECORDS AND IMAGE ASSOCIATIONS Each line of trace/actions.jsonl describes a recorded action, with its identifier, sequence in- dex, action kind, video-relative timestamp, and caption. Coordinates and target information are retained when available. The following excerpt reproduces selected fields from the first action in the hard app-f016 episode used in Appendix F: "sequence_index": 0, "action_id": "act_0JYEZK480Ta_81AR", "kind": "click", "media_time_ms": 3908, "caption": "Left Click push button - Calendar", "target_label": "push button - Calendar" When present, evidence/manifest.json and evidence/image actions.jsonl provide the image index and its input-event and trajectory associations. In this episode, the first image-index record references the same actionid, with frametimems = 3908 and actiondeltams = 0. These fields let the agent locate the corresponding event frame and nearby video observations. A nonzero actiondeltams indicates an approximate temporal match. An event frame captures the interface near the input event; verifying its effect may require a subsequent observation. B.3EVIDENCE AVAILABILITY AND COUNTS The schema describes the files available within an episode, rather than guaranteeing every modal- ity for every demonstration. Target metadata in an action record does not imply that a complete accessibility-tree sequence is included. In Linux mode, the continuous tree sequence remains in the Linux evidence directory, while the Mac export can preserve snapshots received with events. The induction agent can inspect only the evidence provided within its exposed input directory. The manifest distinguishes action records, input events, recorded keyframes, and exported evidence images. Table 1 reports mean counts of action records and exported images per episode. The hard app-f016 episode, for example, contains 37 action records and 23 evidence images, alongside 176 input events and 29 recorded keyframes. Action and image counts differ because they use separate grouping rules. Image export pairs key presses with releases, groups double clicks, and merges adjacent typing evidence into one image with the last frame and a combined text annotation. Scroll records do not independently generate images, while action preprocessing merges adjacent typing and scroll records. The converter sets counts.keyframes from the recordingâs keyframe count, falling back to the number of loaded keyframe records; counts.evidence images counts images copied into the main evidence collection. Exporting all available keyframe screenshots is a separate option that creates evidence/all keyframes/ and its own manifest. Recorded keyframes thus reflect the capture policies in Appendix A.2; exported images reflect subsequent evidence processing. 16 Preprint CEVIDENCE-EXPLORATION TOOL DEFINITIONS Table 5: Implemented induction tools. NexAU additionally provides LoadSkill for loading the bundled authoring policy. ToolCapability listdirectoryList files and subdirectories within an exposed directory. globFind files matching a glob pattern within the exposed input or output directory. read fileRead text and structured evidence with line numbers and pagination. read visualevidenceReturn an ordered batch of 1â20 images labeled with their source paths. get videometadataInspect video duration, dimensions, frame rate, codec, and file size. watch videoSample timestamped frames over a selected interval, subject to a frame limit. extractvideoframesRequest frames at selected millisecond timestamps. writefileCreate or replace a complete output file. finalizeextractionCheck output structure and tool-use requirements, then write the extraction manifest on success. The input directory is exposed as . and generated files as output/. File tools resolve paths within these roots; writes are limited to the output directory and cannot replace runtime-owned manifests or logs. Task interpretation, action selection, and comparisons between observations are performed by the agent using the returned evidence. The following descriptions summarize the registered tools and their principal arguments. File discovery and reading. list directory takes dirpath; glob takes a pattern and op- tional dirpath. Both support filtering options. readfile takes filepath, with optional zero- based line offset and limit. It handles text, JSON, JSONL, YAML, and Markdown as text, returns numbered lines and continuation guidance, and directs image and video requests to the cor- responding visual tools. A response is limited to 20,000 characters of numbered content; individual long lines are truncated. This interface provides access to task metadata, action traces, input logs, image indices, and any exported structured UI artifacts. Image inspection. read visualevidence takes an ordered filepaths list containing 1â20 distinct images. Optional imagedetail and imagemaxsize control detail and resizing. Re- turned images preserve the requested order and have source-path labels, allowing the agent to com- pare selected states while retaining their file identities. Video inspection. getvideometadata takes filepath and reports metadata obtained with ffprobe. watchvideo takes the video path and optional startms, endms, intervalms, maxframes, and framewidth. It returns sampled images; defaults are a 5,000-ms interval, at most 12 frames, and a width of 1,280 pixels. The interval must be at least 100 ms and each call can request at most 20 frames. extractvideoframes instead takes a list of 1â20 timestampsms and an optional framewidth. Both use ffmpeg and return timestamp labels. Near the video end, the requested time can be adjusted backward to obtain a decodable frame; the label reports the sam- pling time used and the original request when adjusted. Finalization checks that metadata inspection preceded sampling for the same video path. Writing and finalization. write file takes filepath and the complete content, creating or replacing a file. finalizeextraction takes no required arguments. It requires at least one skill, a recorded read of an input task.json, valid skill frontmatter and a nonempty body, a writefile call and a subsequent readfile call for every final skill, and the video-metadata ordering described above. The skill name must use lowercase hyphen-case and match its folder; its description must 17 Preprint be nonempty. Heuristic checks for fixed coordinates, local paths, and timestamps produce warnings rather than semantic correctness guarantees. C.1EXAMPLE EXPLORATION TRACE An illustrative sequence using these tools is: 1. load skill-creator, locate the authoritative task.json, and read it with read file; 2. discover evidence files with listdirectory and glob, then read relevant traces and image indices; 3. inspect representative initial, critical, and final-state images with readvisualevidence; 4. if a transition remains unclear, read the videoâs metadata and use watchvideo or extract videoframes to examine it; 5. write a skill with write file, inspect further evidence if needed, and revise the draft; and 6. read the final SKILL.md after its last write and call finalizeextraction. This is an example, not a prescribed sequence of all tool calls. The system prompt sets the task- first and representative-image requirements; the agent chooses the detailed exploration and revision sequence. On successful finalization, the runtime manifest records skill hashes, inspected source paths, validation results, and provenance. Separate run logs capture tool activity and available usage information. These records provide run-level provenance without asserting a source citation for every generated skill statement. DPROMPTS D.1INDUCTION SYSTEM PROMPT The induction system prompt sets the task, evidence-access boundaries, and completion require- ments. Its instructions are reproduced below. Induction System Prompt You extract reusable NexAU skills from one recorded computer-use benchmark episode. ⢠Stay within the exposed input and output roots. Never seek grading rubrics, judge output, reference answers, or files outside them. ⢠First load skill-creator; it is the sole authority for Skill boundaries and content. Do not restate, shorten, or weaken its authoring rules. ⢠Read the authoritative task.json before interpreting the trajectory. Treat traces and recordings as fallible evidence; recording completion is not task success. ⢠When indexed evidence images exist, inspect representative initial, critical, and final-state frames. Before sampling any video, call get videometadata for that exact path. ⢠Write generated Skill folders only under the output root. Do not create or edit runtime-owned extraction-manifest.json, run-events.jsonl, run-metadata.json, or run-summary.json. ⢠Read each final SKILL.md after its last write, then call finalizeextraction. Correct every reported error and retry. Only successful finalization completes the task. D.2NEXAU AGENT CONFIGURATION AND COMPLETION The launcher creates one NexAU agent with the system prompt, registered tools, and bundled skill-creator policy. Its initial message directs the agent to load that policy, read the task, cre- ate skills under the output root, read back each final skill, and finalize. The implementation uses 18 Preprint structured tool calls; it does not prescribe a separate controller state or a numerical information-gain objective. Evidence inspection and skill authoring occur in the same agent conversation. The model and endpoint are supplied at launch, with configurable iteration, context, output, tempera- ture, and reasoning settings. Our experiment settings are reported in Appendix E.5. Provider-request retries in the supplied implementation are configured with 30 attempts, a maximum backoff of 30 seconds, and a 900-second request timeout. The prompt instructs the agent to correct finalization errors and retry. finalize extraction re- turns diagnostics on failure and atomically writes extraction-manifest.json on success. The launcher checks that finalization succeeded after the agent returns; an ordinary text response alone leaves the run incomplete. The runtime also writes configuration provenance, event logs, and a sum- mary of status, timing, tool activity, and available token usage. Structural validation and recorded call ordering are enforced by code; the authoring policyâs requirements for evidence grounding, workflow completeness, and appropriate generalization remain instructions to the model. D.3PIPELINE PROMPT The Pipeline Prompt baseline changes only the induction system prompt, keeping the model, in- put evidence, tools, skill-authoring policy, and output format the same as Work2Skill. It expands the concise bootstrap prompt into an explicit workflow for reading the task, inventorying arti- facts, inspecting structured evidence and images, sampling video, and authoring and validating skills. Both conditions load the same skill-creator policy, read back final skill files, and call finalize extraction. The detailed prompt prescribes an inspection order and frame-selection rules, including a focused initial image batch and further inspection when ambiguity remains. One possible explanation for the prompt-ablation results (Section 4.5) is that a predefined exploration route may narrow the range of evidence the model considers. The baseline prompt is reproduced below, with formatting adapted for readability. Pipeline Prompt You extract reusable NexAU skills from recorded computer-use benchmark artifacts. Boundaries ⢠The tools expose one logical input root and one logical output root. Treat those boundaries as absolute. ⢠Never seek grading rubrics, judge output, reference answers, or files outside the exposed roots. ⢠The runtime owns extraction-manifest.json, run-events.jsonl, run-metadata.json, and run-summary.json. Never create, edit, or overwrite those files. ⢠A normal text response does not complete the task. Only a successful finalizeextraction call completes it. ⢠The task instruction is authoritative. Recording status only means recording finished; it does not prove task success. ⢠Read the authoritative task instruction before interpreting the trajectory. Preserve its goals, quantities, exclusions, ordering constraints, and requested scope while parameterizing episode-specific values. ⢠A demonstrated state change teaches interface mechanics but does not authorize a broader goal or an additional persistent mutation. ⢠Treat screenshots, action logs, key logs, summaries, and video as fallible evidence that must be reconciled. ⢠Before the first watch video or extractvideoframes call for any video path, call getvideometadata for that exact path. This ordering is irreversible within a run: calling metadata after sampling cannot repair it, and runtime event logs must never be edited. 19 Preprint Pipeline Prompt (continued) Image Evidence ⢠Files under an episode's evidence/ directory are event-aligned evidence frames, not arbitrary screenshots. They are decoded from the episode video at frame timems, which is derived from a grouped key or pointer event after applying frameoffsetms. ⢠Read evidence/manifest.json or evidence/imageactions.jsonl before selecting images. These indexes connect each image to its event kind, event time or interval, pointer points, and the nearest trajectory actionid and caption. A nonzero actiondeltams means the event and trajectory action were only approximately aligned. ⢠The image annotation is generated evidence metadata, not part of the recorded application UI. Its top banner describes the input event and matched trajectory action. Click and double-click frames mark the pointer target; drag frames show the recorded path; typing frames are taken at the end of the grouped typing interval and describe the recorded keys. ⢠Use these full-resolution frames to identify visible labels and semantic controls, verify which item or field was targeted, read entered or selected values, understand drag endpoints, and compare UI state across important actions. They are especially useful where coarse video sampling makes text or state changes ambiguous. ⢠An event frame shows the UI at or near the input event. Do not assume it proves the action's resulting state. Verify cause and effect by comparing the next relevant evidence frame or by extracting a video frame shortly after the event. Because images are event-triggered, the final successful state may require inspection near the end of the video. ⢠Pointer coordinates, timestamps, generated crosshairs or paths, and one episode's action caption are trace evidence only. Convert them into semantic targets and observable state; never copy them into a reusable skill as fixed instructions. ⢠If an episode contains evidence images, inspect 1â20 representative images together with read visualevidence. Cover the initial state, critical choices or inputs, important transitions, and the final observable outcome. watchvideo samples do not replace this full-resolution inspection. Workflow 1. Locate and read the authoritative task.json before interpreting the trajectory, then inventory the input recursively with listdirectory and glob. Pass dirpath: "." for the logical input root; use dirpath: "output" for generated files. 2. Read textual and structured artifacts with readfile. The same tool handles text, JSON, JSONL, YAML, and Markdown. 3. Read the evidence index, select representative initial, critical, and final event frames, and inspect them with read visualevidence. Prefer one focused batch and inspect more only when ambiguity remains. A one-item filepaths list reads a single frame; the tool accepts up to 20 paths. 4. For video, call getvideometadata for the exact video path before any sampling. Only then use watchvideo for coarse or local interval sampling and extractvideoframes at exact action timestamps when details matter. 5. Load skill-creator before authoring output. 6. Infer reusable procedures, semantic target-selection rules, observable completion checks, and supported recovery behavior. Separate these from episode-specific values and incidental human actions. Keep phases in one composite Skill when they share state, task-derived data, requested counts, or required order. Split only independently invocable capabilities. 7. Write each skill as <skill-name>/SKILL.md under the output root. Add references or scripts only when they directly improve reuse. 8. Before finalizing, read every final SKILL.md back with readfile after its last write and self-review it: ⢠SKILL.md starts with YAML frontmatter containing a lowercase hyphen-case name matching its folder and a non-empty description. ⢠The body is non-empty, reusable, and uses semantic targets rather than fixed coordinates. 20 Preprint Pipeline Prompt (continued) ⢠Every explicit task goal, constraint, requested quantity, selection rule, scope, and required ordering relationship needed at runtime is present in the Skill. ⢠Every persistent state change instructed by the skill is authorized by the task. ⢠It contains no machine paths, episode-specific timestamps, action indexes, session IDs, or instance values presented as general procedure. ⢠Any material uncertainty is expressed in the Skill as an observable conditional check or limitation. 9. Call finalize extraction. It confirms that task.json was read, every final Skill was read back after its last write, video metadata ordering was respected, and all final Skill files are valid. It writes the runtime-owned manifest. If it reports any problem, correct the output and call it again. Do not finish with an ordinary text response. Quality Rules ⢠Do not encode fixed screen coordinates, action sequence indexes, timestamps, local machine paths, session IDs, or one recording's dates, names, booking IDs, and other instance values as general procedure. ⢠Prefer semantic UI roles, visible labels, relative state, and task-derived selection criteria. ⢠Omit actions that appear to be retries, misclicks, idle behavior, or exploratory scrolling unless they reveal a reusable recovery strategy. ⢠State only what the instruction or observable artifacts support. Express material uncertainty as an observable check instead of inventing hidden success criteria. ⢠Use the minimum sufficient operational detail. The Skill must remain executable without the source recording or extraction manifest. D.4SKILL-CREATOR Skill-Creator --- name: skill-creator description: Create self-contained, operationally complete skills from inspected computer-use benchmark evidence. Use when an extractor must convert task instructions, action traces, screenshots, and video observations into portable skill folders without copying episode-specific details. --- Create Reusable Skills Convert observed task behavior into instructions that another agent can execute with fresh task inputs. Describe stable decisions, semantic targets, visible state, completion checks, and supported recovery rather than replaying the recorded trajectory. This Skill is the authoring policy for generated Skills; do not replace its requirements with a shorter summary. Runtime Contract The execution agent receives the generated Skill and its bundled resources, not the source recording, action trace, or extraction manifest. Put every correctness-critical instruction in SKILL.md; supporting references must not be required to reconstruct an omitted workflow, formula, selection rule, quantity, ordering constraint, stop condition, or completion check. For a graphical computer-use task, preserve the interaction modality. Instruct the agent to make task-state changes through visible application controls and verify them in the visible application. Never substitute an API, database edit, script, browser console, developer tool, or other hidden state mutation unless the authoritative task explicitly requires that mechanism. 21 Preprint Skill-Creator (continued) Length is not a quality target. Remove generic explanation and trace provenance, not operational invariants. If removing a detail could change the selected target, computed value, action order, affected quantity, visible success state, or recovery behavior, retain it. Define the Skill Boundary Create one Skill for each coherent reusable capability. Keep phases together when they exchange task-derived data, share application state or counts, or must execute in a particular order. Split Skills only when each capability is independently invocable and useful; do not fragment one workflow merely to make individual files shorter. Choose a short lowercase hyphen-case name under 64 characters. Use the same name for the output folder and the frontmatter name. Separate Procedure from Trace Treat the task instruction as the goal and the recording as fallible evidence. Read the task first, then reconcile task data, action logs, evidence images, and video before stating the reusable procedure. Recording completion alone does not prove task success. A recorded state change demonstrates interface mechanics but does not authorize an extra persistent mutation. Preserve only state changes required by the task, and retain: ⢠semantic application, page, dialog, field, item, and control names; ⢠rules for choosing the correct target from changing task inputs; ⢠exact formulas, operands, units, basis, thresholds, rounding rules, and where each computed result is used; ⢠requested quantities and scope, including per-item repetition and count-based stop conditions; ⢠dependencies, required order, and data or state passed between phases; ⢠visible checks after critical mutations and a final observable completion check, including refresh or reopen checks when persistence matters; ⢠evidence-supported recovery from loading delays, stale searches, hidden controls, misclicks, dialogs, validation errors, and partial completion. Remove screen coordinates, action indexes, timestamps, session identifiers, local machine paths, and retries or idle actions that do not teach recovery. Replace one episode's names, dates, IDs, routes, values, filenames, and messages with task-derived parameters when they are inputs rather than reusable constants. Parameterize episode-specific values without abstracting away the formula, threshold, relationship between inputs, requested count, ordering constraint, selection rule, or visible verification that makes the procedure correct. Do not promote one observed layout, value, or incidental action into a universal requirement. When evidence is incomplete, state an observable conditional check or supported limitation in the Skill. Never invent a hidden rubric or success condition. Write the Skill Create <skill-name>/SKILL.md with write file. Start with exactly this frontmatter shape: --- name: skill-name description: Explain what the skill does and the situations that should trigger it. --- Write the body in clear imperative language. Include the complete workflow, selection and calculation rules, scope and count constraints, phase handoffs, visible completion checks, and only recovery guidance supported by evidence. Avoid introductory material a capable execution agent already knows, but never remove a step, dependency, decision, stop condition, or recovery check that can affect correctness. 22 Preprint Skill-Creator (continued) Use semantic UI targets instead of fixed coordinates. When labels or layouts may vary, give a stable visible role or relationship that identifies the target without guessing. For persistent changes, verify the resulting visible state; refresh, revisit, or reopen the relevant view when the observed workflow shows that this is necessary to distinguish a saved result from a transient one. Add references/ only for genuinely supporting knowledge, never as a place to hide required runtime instructions. Add scripts/ or assets/ only when they are essential to reuse and their correctness can be established from the available evidence. Do not initialize templates, run packaging scripts, or create archives. Self-Review Before finalization, read every final SKILL.md back with read file after its last write and verify: ⢠the folder name and frontmatter name match, the description identifies the right trigger, and the body is non-empty; ⢠an execution agent can complete the workflow without the source recording, trace, extraction manifest, or unstated episode knowledge; ⢠graphical tasks make state changes through visible UI actions and do not offer backend or developer shortcuts unless the task explicitly requires one; ⢠every explicit goal, formula, operand, unit, quantity, scope, selection rule, ordering relationship, phase handoff, and count-based stop condition needed at runtime is present; ⢠semantic targets replace coordinates and incidental layout; ⢠critical mutations and the final result have observable UI checks, including persistence checks where required; ⢠every instructed persistent mutation is authorized by the task; ⢠supported failure modes retain their recovery and stop conditions; ⢠episode-specific values are parameterized without erasing the relationships that govern their use; ⢠material uncertainty is an observable conditional check or limitation, not an invented action or hidden success criterion; ⢠every bundled file is referenced, necessary, and contains no core instruction that belongs in SKILL.md. Call finalize extraction. The runtime writes the extraction manifest. Treat the finalizer result as authoritative: correct every reported issue and call it again. Do not finish with an ordinary text response. D.5DOWNSTREAM EXECUTION PROMPT We append the following shared wrapper to the executorâs base system prompt for Work2Skill, ShowUI-Aloha, and D2T. Only the guidance text changes between methods, as detailed in Table 6; No Skill adds no guidance block. The wrapper directs the executor to adapt the guidance to the current task and screen and to verify the result in the live environment. Shared Downstream Guidance Wrapper <taskskills> The following reusable skills are advisory guidance for this task. Adapt them to the live screen and current task values. The userâs task instruction remains authoritative. Do not assume recorded coordinates, timestamps, identities, instance values, or completion state. Verify the requested result in the live environment before finishing. [guidance text] </taskskills> 23 Preprint Table 6: Method-specific contents inserted into the shared <taskskills> wrapper at the end of the system prompt. No Skill adds no block. MethodRetained contentText format No SkillNoneNo appended block. Work2SkillComplete task-matched SKILL.md files, including their frontmatter and instructions. Each file follows a skill-directory heading. Multiple files are concatenated in the task mapâs order, separated by blank lines. ShowUI-AlohaEach trajectory stepâs caption.action text. A task heading followed by Step 1: ..., Step 2: ..., and subsequent actions in trajectory order. D2TTutorial prose after removing frontmatter, image-containing lines, and link destinations. Original section and step headings and remaining prose; ordinary hyperlink labels are preserved. EBENCHMARK TASKS AND EVALUATION Task screening and demonstration review. We first exclude infeasible tasks for which a valid human demonstration cannot be completed. We then review the human recordings and discard demonstrations that are incomplete, omit key operations, or fail to complete the task correctly. Only tasks and their associated demonstrations that pass this screening and review are retained. E.1MYPCBENCH TASKS AND JITTER PERTURBATIONS We use the 64-task MyPCBench Bounded Action task set. Each task is paired with a recorded human demonstration and its exported multimodal evidence for skill induction. Perturbation construction. We construct one Jitter variant for each original task, yielding 64 sourceâvariant pairs. GPT-5.6-sol assists in generating the revised task instructions and correspond- ing evaluation criteria, followed by human verification. The perturbations modify task requirements within the original application context: target entities and source materials; numerical and temporal parameters; selection and ranking rules; and operation constraints or required final states. Table 7 gives representative examples. Evaluation criteria and skill reuse. All 64 task instructions are revised. Across the 366 paired rubric items, 331 criteria are updated to match the perturbed requirements and 35 generic criteria are retained. All rubric weights and evaluator types remain unchanged. Each Jitter task retains the skill-file mapping of its paired original task, so the evaluation measures reuse of the existing skill guidance on a changed task instance. Figure 7 illustrates paired changes to a SprintBoard task and its evaluation criteria. Table 7: Representative changes in the paired MyPCBench Jitter tasks. Descriptions summarize the changed requirements. ApplicationOriginal requirementsJitter requirements Dinoco AirlinesBook an AVPâJFK flight for next Friday evening or night with a left-side seat. Book an AVPâPHL flight for next Tuesday morning with a right-side seat. OddsMarketSelect the three largest positions by share count; break ties by higher cost basis, then alphabetically. Select the two smallest positions by share count; break ties by lower cost basis, then alphabetically. SpeedTaxAdvance the in-progress current-year re- turn to the review step without submitting it. Audit the filed 2024 return read-only and report refunds, confirmation number, and accepted status. 24 Preprint Original Task instructions: PopopenmyDiversityDayPlanning2026boardon SprintBoardanddragacardoutoftheTodocolumn intoInProgressâjustonecard,whicheveryouthinkis themostsensiblenextstepforDiversityDayprep. Jitter Task instructions: OpenMovieMondayPrograminSprintBoardanddrag exactlyonecardfromInProgresstoInReview.Verify thechosencardremainsinInReviewafterthedrag. â˘AgentopensSprintBoardandnavigatestothe DiversityDayPlanning2026project â˘AgentopensSprintBoardandnavigatestothe MovieMondayProgramproject Rubric Rubric â˘AgentpicksacardfromtheInProgresscolumnand performsaleft_click_draggesturetoInReview â˘AgentconfirmstheSprintBoardcardappearsin theInReviewcolumnafterthedragreleaseonthe MovieMondayProgramboard â˘AgentlocatesaspecifictaskcardsittingintheIn ProgresscolumnoftheMovieMondayProgram SprintBoardbeforeinitiatingthedragtoInReview â˘AgentpicksacardfromtheTodocolumnand performsaleft_click_draggesturetoInProgress â˘AgentconfirmstheSprintBoardcardappearsinthe InProgresscolumnafterthedragreleaseonthe DiversityDayPlanning2026board â˘AgentlocatesaspecifictaskcardsittingintheTodo columnoftheDiversityDayPlanning2026 SprintBoardbeforeinitiatingthedragtoInProgress Figure 7: A paired SprintBoard task and its evaluation rubric before (left) and after (right) Jitter perturbation. The variant changes the target board and the required card transition, with rubric items revised to match the new requirements. E.2OSWORLD-VERIFIED TASK DISTRIBUTION The curated OSWorld-Verified subset contains 286 tasks across ten categories. Figure 8 summarizes the category distribution. Percentages are calculated over the 286 retained tasks. The complete task-ID list appears in Table 12 in Appendix H. 286 tasks CategoryTasksShare (%) Cross-application 6924.13 Chrome 4114.34 LibreOffice Calc 3913.64 LibreOffice Impress 3311.54 Operating system 227.69 LibreOffice Writer 217.34 VS Code 217.34 GIMP155.24 Thunderbird 155.24 VLC103.50 Figure 8: Category distribution of the 286 retained OSWorld-Verified tasks. The legend gives the task count and percentage for each category. E.3EXECUTION HARNESS AND EVALUATOR CONFIGURATION OSWorld execution harness. We use a tool-calling desktop agent connected to OSWorldâs official DesktopEnv. Each task is initialized from its benchmark configuration. Given the task instruction, the agent uses the current screenshot and action history to select its next GUI actions, with guidance 25 Preprint appended as described in Appendix D.5. Its tools support screenshots, pointer movement, clicks, drags, scrolling, text entry, key presses, and waiting. A desktop adapter maps screenshot-relative pixel coordinates to the VM and executes GUI actions through PyAutoGUI, returning post-action screenshots and textual results. Each execution step is one model iteration and may contain multiple tool calls. The loop stops on a terminal decision or at the step limit specified in Appendix E.5. Evaluation and metrics. For MyPCBench, we use the official environment setup and evaluator. For OSWorld-Verified, we reuse the official task-specific evaluators to score the resulting environ- ment states. Evaluators run after the agent terminates or reaches the maximum action budget. For partial MyPCBench rubrics, the task score is the fraction or weighted fraction of satisfied items ac- cording to the official implementation. Full score requires all rubric conditions. OSWorld-Verified evaluators can return partial scores. For both benchmarks, Score averages the evaluator outputs and scales the result to 0â100, while SR counts only tasks receiving full credit. This SR corresponds to MyPCBenchâs perfect rate; OSWorldâs official mean-score aggregation, called âsuccess rateâ in its reporting script, is reported as Score in our tables. We store evaluator logs and final-state artifacts for every rollout. Let g i â [0, 1] denote the normalized evaluator output for task i, where 1 indicates full credit. For a task set D, we compute the mean Score and full-credit success rate SR (%) as follows: Score = 100 |D| X iâD g i ,SR = 100 |D| X iâD 1[g i = 1] .(1) E.4BASELINE ADAPTATION FOR FAIR COMPARISON No Skill (also labeled Baseline) provides no additional guidance. We evaluate ShowUI-Aloha and D2T as sources of guidance for the same downstream execution agent. For ShowUI-Aloha, we retain each trajectory stepâs caption.action field and concatenate the actions in their original order; observation, reasoning, and expectation fields are excluded. For D2T, we read the generated tutorialâs index.md, remove frontmatter and image-containing lines, and strip link destinations and bare URLs while preserving ordinary hyperlink labels. The remaining tutorial headings and prose retain their original order. All retained text is supplied through the same wrapper as our skills (Appendix D.5), with the method-specific contents summarized in Table 6. For a given execution model, all methods share the base prompt, observations, action interface, step limit, termination policy, and evaluator. On MyPCBench, this shared action interface includes both GUI actions and Bash commands. When their required modalities are available, they use the same source demonstration. The reported results measure how effectively these adapted guidance outputs support task execution under the same downstream conditions. E.5MODEL PAIRING AND RUNTIME SETTINGS The default pairing uses the same model for skill induction and downstream execution: each of the four execution models uses its own induced skills. The cross-model experiment explicitly varies this pairing (Section 4.4). Tables 8 and 9 report our settings for skill induction and downstream execution, respectively. The induction limit counts agent iterations; the execution limit counts steps. The execution history and screenshot windows are configured separately from token limits. Table 8: Our skill-induction settings. SettingValue Reasoning efforthigh Temperature0 Context limit (tokens)200,000 Output limit per model call (tokens)16,000 Induction iteration limit100 Table 9: Our downstream-execution settings. SettingValue Reasoning efforthigh Temperature0 Output limit per model call (tokens)16,000 History window (turns)100 Screenshot window (images)3 Execution step limit100 26 Preprint Execution context management. When a downstream model request exceeds the context limit, the system retains the task instruction, skill guidance, and most recent screenshot, clears the histori- cal interaction messages, and retries the request. This shortens the context while preserving the task requirements and visual evidence of the current interface. E.6EVIDENCE-MODALITY ABLATION SETTINGS The ablations in Section 4.5 vary the recording evidence available to the induction agent. The task instruction remains available in every condition. Table 10 lists the retained evidence and visual tools. We remove only tools that have no applicable input under a condition. File discovery, text reading, skill authoring, and finalization use the same tools described in Appendix C. Table 10: Evidence and visual tools available for skill induction in each modality condition. Task instructions and common file-reading and skill-authoring tools remain available. UI annotations denote metadata and descriptions derived from each actionâs target element in the AX tree. SettingRetained recording evidenceVisual tools Full EvidenceActions with UI annotations, KeyLogs, indexed images, videoImage and video w/o Indexed ImagesActions with UI annotations, KeyLogs, videoVideo w/o UI AnnotationsActions with generic captions, KeyLogs, images with UI target annotations removed, video Image and video w/o KeyLogsActions with UI annotations, event frames without added over- lays, video Image and video Video-onlyRaw videoVideo UI-Grounded Action Text Only Ordered text describing recorded actions and their target UI ele- ments None In Table 10, the image tool is readvisualevidence; video tools are getvideometadata, watchvideo, and extractvideoframes. Removing UI annotations or KeyLogs does not re- quire removing readfile, which still serves the remaining textual evidence and task instruction. Image and video access. w/o Indexed Images removes the exported evidence/ directory and the image-reading tool, while retaining the video and all video tools. The agent can therefore still inspect video frames; this condition tests the contribution of indexed, event-aligned images when video remains available. Video-only exposes only task.json and video.mp4. Video metadata and frame timestamps remain available through the video tools; no action trace or precomputed image index is supplied. Removing derived annotations. The condition labeled w/o UI Annotations corresponds to the implementationâs without dom protocol. It removes AX-derived target-element fields from actions and replaces the corresponding target descriptions with generic action captions. Image indices are sanitized, and the top annotation bar is replaced with raw-video pixels from the same timestamp, preserving click and drag highlights. w/o KeyLogs removes trace/keylog.jsonl and keylog-derived image metadata, including input- event labels, raw messages, and source-image references. Indexed frames are re-extracted from raw video at their original timestamps, removing all added image overlays, including the top caption bar and click/drag highlights. The action trace and its AX-derived target information remain unchanged, as do the action captions in the image index. Thus, this condition measures the joint effect of remov- ing the raw log and the annotated image presentation; it does not isolate the log fileâs contribution or remove all keyboard information from the retained actions. UI-grounded action text only. This condition retains only the task instruction and ordered action captions in task.json and trace/actions.jsonl, respectively. The captions combine recorded actions with available target UI descriptions, such as âclick the Search button.â Target descriptions 27 Preprint can include control names and roles from the corresponding target elements in the AX tree. Each action row retains only its original caption; coordinates, timestamps, action IDs, structured target metadata, and other action metadata are removed. Raw keylogs, images, and video are absent, and all image and video tools are disabled. Thus, UI target semantics are preserved in text rather than supplied as a full AX Tree. This condition evaluates skill induction from these descriptions; ShowUI-Aloha instead supplies action semantics directly as downstream guidance. Prompt adaptation. The five ablations share a prompt instructing the agent to inspect the available evidence, followed by a condition-specific statement listing the exposed modalities. This adapts the full-evidence promptâs requirement to inspect representative initial, critical, and final-state images to the available inputs. Task-first reading, the skill-authoring policy, output validation, and finalization requirements are retained. FCASE STUDY AND SKILL INDUCTION EXAMPLE F.1TASK AND INDUCTION TRACE We present the generated skill for MyPCBench Bounded Action task hard app-f016, a TableFind reservation task also used in the main-text case study (Section 4.6). The task instruction is: Book me Cotogna for ten people for the Dundies after-party next Friday. While youâre in TableFind, pull up three other Scranton-area restaurants as a fallback plan in case Cotogna is full, and favorite the top two. Once thatâs done, confirm the Cotogna reservation with a note that itâs the Dundies after-party so they can prep the vibe. The recorded induction run uses GPT-5.5 with high reasoning effort and makes 30 tool calls. It reads the task, episode metadata, summary, evidence indices, action trace, and key log; inspects 22 distinct evidence images in two batches of 14 and eight; and samples the 156.47-second video in two calls to watchvideo, after reading its metadata. The video calls return 13 and eight frames, respectively, with one shared timestamp. The agent writes a draft, reads it back, revises it, and reads the final skill before successful finalize extraction. This record documents the induction output; we do not assign a downstream execution score to this listing. F.2DETAILED EXECUTION ANALYSIS Figure 6 shows GPT-5.5âs execution trajectories for the case in Section 4.6. Table 11 reports all eight rubric outcomes under ShowUI-Aloha, D2T, and Work2Skill guidance. Task dependencies. The instruction combines a primary reservation with a fallback workflow: inspect three other Scranton-area restaurants and favorite the top two before confirming Cotogna with the requested note. Completing this workflow requires carrying the date and party size across searches, distinguishing the primary restaurant from alternatives, and returning to the primary book- ing after handling the fallbacks. The evaluation checks both the reservation itself and these addi- tional requirements, including persistence after a refresh. Where the baselines fall short. Both baselines pass R1, R2, R6, and R8: they configure the book- ing, include the note, obtain confirmation, and verify the final reservation details. Their trajectories reach alternative-restaurant views, but the rubric outcomes do not establish that all three alternatives and their detail pages were inspected (R3âR4). Saving the top two alternatives is also incomplete (R5). Both fail the separate persistence check (R7), despite passing the check of final details (R8). A visible confirmation and a correct reservation display therefore cover only part of the required workflow. Each baseline receives 55/100. 28 Preprint Table 11: Rubric-level outcomes for the TableFind case in Figure 6. Requirement descriptions summarize R1âR8; Aloha denotes the adapted ShowUI-Aloha baseline. Scores are the reported evaluator outputs. Item RequirementAlohaD2TWork2Skill R1Set the requested date and a table for ten.PassPassPass R2Book Cotogna with the requested note.PassPassPass R3Open three alternative restaurants.FailFailPass R4View each alternativeâs detail page.FailFailPass R5Save the top two alternatives.FailFailPass R6Obtain a reservation confirmation number.PassPassPass R7Check persistence after a refresh.FailFailPass R8Verify the correct final details.PassPassPass Evaluator score55/100 55/100100/100 What the successful trajectory preserves. Work2Skill passes all eight items and receives 100/100. Its four additional passed criteria concern alternative-restaurant coverage, saving the pre- ferred options, and persistence after refresh. Figure 6 shows these steps alongside the primary book- ing and confirmation. The comparison pinpoints the benefit in this case: completing the secondary subgoals and verification requirements that remain unfinished in the baseline trajectories. Connection to the generated procedure. The full skill in Appendix F.3 makes several relevant dependencies explicit. It carries booking values across searches, rechecks fields that may reset, excludes the primary restaurant from the fallback set, and selects the first two qualifying alternatives under the active ranking. It also defers final confirmation until the fallback steps are complete and checks both the reservation details and the saved favorites. These instructions organize the task into dependent stages with observable completion checks and explicit reminders of the secondary requirements across the primary reservation and fallback searches. F.3COMPLETE GENERATED SKILL The skill below is reproduced in full, with only Markdown presentation adapted to the appendix format. It carries the date, time, party size, and note across the primary and fallback searches; checks for search-field resets; excludes the primary restaurant from the fallback set; and preserves the ordering of fallback selection, favoriting, and final confirmation. Restaurant names and booking values are taken from the task, while the generated procedure retains the Scranton-area scope. Its final checks use visible reservation and favorite states. 29 Preprint Generated Skill: TableFind Reservation with Fallbacks --- name: tablefind-reservation-with-fallbacks description: Use TableFind to reserve a requested restaurant for a specified party/date/time, identify fallback Scranton-area restaurants, favorite the requested top fallback choices, and verify the confirmed reservation details. --- Use this workflow when a task asks for a TableFind restaurant reservation plus fallback restaurants and favorites. Inputs to carry through the workflow ⢠Primary restaurant name. ⢠Party size. ⢠Reservation date expression or exact date. Resolve relative dates such as ânext Fridayâ before interacting with the calendar; if the task is based on the current TableFind/session date, use that context to choose the next matching weekday. ⢠Reservation time. If the task does not specify a time and TableFind already provides a default time in the search form, leave the default time unchanged rather than inventing a different one. ⢠Reservation note or special-request text. ⢠Fallback scope, count, and favorite rule, such as âthree other Scranton-area restaurantsâ and âfavorite the top two.â Keep the same date, time, and party size for fallback searches unless the task explicitly gives different fallback criteria. Fallback restaurants must be different from the primary restaurant. Set the primary search criteria 1. Open TableFind and use the visible Search page or TableFind logo/home link to reach the main search form. 2. Set the Date field to the task date. Use the calendar picker and select the resolved day; verify the date field updates to the intended date. 3. Leave or set the Time field to the task time. If no time is specified, keep TableFindâs visible default time. 4. Open the Party Size combo box and choose the requested party size, e.g. the menu item labeled with the required number of people/guests. Verify the party-size field shows the requested count. 5. Click the Restaurant or Cuisine entry, type the primary restaurant name, and click Search. 6. On the results page, verify the heading matches the requested party size, date, and time, and select the result whose visible name exactly matches the primary restaurant. Check and prepare the primary reservation without losing task details 1. On the primary restaurant detail page, verify the restaurant name, cuisine/location summary, booking panel date, party size, and available time slots are visible. 2. Select the requested time slot if it is not already selected. The selected summary should show the intended date and time and indicate remaining tables, not sold out or waitlist. If the exact requested/default time is unavailable, do not confirm an alternate unless the task authorizes alternatives. 3. Enter the taskâs note in the Special requests field exactly enough to communicate the requested purpose or constraint. Verify the note text is visible in the field. 4. If the task says to gather/favorite fallbacks before confirming the primary reservation, do not click the final Confirm Reservation yet. You may leave the primary detail page and return later, carrying forward the restaurant name, date, time, party size, and note. Pull up fallback restaurants and favorite the top choices 1. Navigate back to the TableFind Search/results view using the TableFind logo, Search link, or Back to search control. 30 Preprint Generated Skill: TableFind Reservation with Fallbacks (continued) 2. Clear the primary restaurant-name search if necessary so that the page shows general Scranton-area restaurant results for the same date, time, and party size. Keep or choose a ranking such as Sort: Top Rated when the task asks for âtopâ restaurants and no other ranking is specified. 3. Verify the results heading still reflects the intended date, time, and party size. If the page has reverted to default values, reset the date, time, and party size before relying on the fallback list. 4. Starting from the top of the ranked Scranton-area results, identify fallback restaurants that are not the primary restaurant. Count only distinct restaurant cards/pages with Scranton-area neighborhood or address information. 5. Pull up the requested number of fallback restaurants by opening each qualifying restaurantâs detail page or otherwise bringing its card/details into view. For each one, confirm it is a different Scranton-area restaurant and that it has availability for the carried date/time/party criteria or a visible set of usable time slots for those criteria. Return to the results list after inspecting each fallback. 6. Favorite the requested top count according to the active ranking. For âfavorite the top two,â favorite the first two qualifying fallback restaurants in the ranked list, not arbitrary later restaurants. Use each restaurant cardâs heart / Save to favorites toggle. 7. After each favorite action, verify visible confirmation: the heart becomes filled/active or the card text changes to âSaved to Favorites.â Do not toggle an already-saved heart off; if it is already active, count it as favorited and leave it active. Confirm the primary reservation 1. Return to the primary restaurant detail page using Search, the primary restaurant name, browser history, or a saved/open page. Re-check the restaurant name, date, time, party size, and special-request note. 2. Click Complete reservation only when the booking panel shows the intended selected date/time and party size and the special-request note is present. 3. In the confirmation dialog, compare every visible row with the task requirements: ⢠Restaurant: primary restaurant name. ⢠Date & time: resolved task date and intended time. ⢠Party size: requested number of guests/people. ⢠Special requests: the requested note text. 4. If any row is wrong, cancel the dialog and correct the visible fields before trying again. If all rows are correct, click Confirm Reservation. Final verification 1. Open My Reservations after confirming. 2. Find the newly created primary reservation. Verify it is marked Confirmed and shows the correct restaurant, date/time, party size, and special-request note. A confirmation code may also be present, but the visible restaurant/date/party/note/status are the essential checks. 3. Verify the requested number of fallback favorites remain active. If still on the results page, check the filled hearts or âSaved to Favoritesâ labels for the top fallback restaurants. If needed, open the Saved page from the header and confirm those fallback restaurants appear there. 4. Stop only after the primary reservation is confirmed and the required top fallback restaurants are visibly favorited. 31 Preprint GTOOL USAGE DURING SKILL INDUCTION Figure 9 summarizes tool usage during skill induction by four models on 64 MyPCBench Bounded Action tasks and the 286-task OSWorld-Verified subset. Each value is the aggregate invocation count divided by the datasetâs task count; the summary row includes all ten tools. GPT-5.5Opus-4.8Qwen-9BQwen-35B 1.001.171.001.00 4.585.168.384.84 2.610.271.910.05 8.025.726.614.81 0.970.480.660.39 1.522.661.451.16 1.250.200.300.02 0.830.060.000.00 1.201.191.971.06 1.001.092.811.00 LoadSkill list_directory glob read_file get_video_metadata read_visual_evidence watch_video extract_video_frames write_file finalize_extraction Total calls / task 22.9718.0025.0814.33 GPT-5.5Opus-4.8Qwen-9BQwen-35B 1.001.021.001.00 3.662.683.883.55 1.320.091.020.05 7.596.554.264.66 0.860.670.660.36 1.421.431.141.16 0.910.390.350.08 0.500.03<0.010.01 1.301.141.051.05 1.071.011.021.02 19.6215.0114.3812.94 0369 (a) MyPCBench Bounded Action 64 tasks (b) OSWorld-286 286 tasks Calls per task Figure 9: Mean tool calls per task during skill induction, using a shared linear color scale. Values are rounded to two decimals; positive values that would round to zero are shown as < 0.01. Horizontal rules separate tool groups; the bold row reports mean total calls per task. Qwen-9B and Qwen-35B denote Qwen3.5-9B and Qwen3.5-35B-A3B, respectively. GPT-5.5 invokes watch video and extractvideoframes most frequently on both datasets, while Qwen3.5-35B-A3B makes the fewest total calls per task (14.33 and 12.94). Invocation counts alone do not establish evidence coverage or skill quality: a call can return multiple images or vary- ing amounts of text. Repeated finalizeextraction calls also do not by themselves identify failed induction runs. When validation fails, the tool returns diagnostics so the agent can address the reported issues and retry. The run can still complete successfully after these intermediate attempts. 32 Preprint HCOMPLETE OSWORLD-VERIFIED TASK LIST Table 12 lists all 286 OSWorld-Verified tasks used for demonstration collection and evaluation. Task IDs are preserved in full and ordered lexicographically, read left to right across each row. Table 12: Complete task IDs for the 286-task OSWorld-Verified subset. No.Task IDNo.Task ID 1 00fa164e-2612-4439-992e-157d019a84362 01b269ae-2111-4a07-81fd-3fcd711993b0 3 02ce9a50-7af2-47ed-8596-af0c230501f84 030eeff7-b492-4218-b312-701ec99e0c 5 0326d92d-d218-48a8-9ca1-981cd6d064c76 035f41ba-6653-43ab-a63-c86d449d62e5 7 04d9aeaf-7bed-4024-bedb-e10e6f00eb7f8 0512b38-d531-4acf-9e7e-0add90816068 9 06ca5602-62ca-47f6-ad4f-da151cde54c10 06fe7178-4491-4589-810f-2e2bc9502122 11 0810415c-bde4-4443-9047-d5f70165a69712 08aced46-45a2-48d7-993b-ed3fb5b32302 13 08c73485-7c6d-4681-999d-919f5c32dcfa14 0a0faba3-5580-44df-965d-f562a99b291c 15 0a211154-fda0-48d0-9274-eaac4ce5486d16 0a2e43bf-b26c-4631-a966-af9dfa12c9e5 17 0b17a146-2934-46c7-8727-73f6b6483e818 0bf05a7d-b28b-44d2-955a-50b41e24012a 19 0cecd4f3-74de-457b-ba94-29ad6b5dafb620 0d8b7de3-e8de-4d86-b9fd-d2dce58a217 21 0e47de2a-32e0-456c-a366-8c607ef7a9d222 0e763496-b6b-4508-a427-fad0b6c3e195 23 0ed39f63-6049-43d4-ba4d-5fa2fe04a95124 0f84bef9-9790-432e-92b7-eece357603fb 25 10a730d5-d414-4b40-b479-684bed1ae52226 12086550-11c0-466b-b367-1d9e75b3910e 27 121ba48f-9e17-48ce-9bc6-a4fb17a7ebba28 1273e544-688f-496b-8d89-3e0f40a0606 29 1334ca3e-f9e3-4db8-9ca7-b4c653be7d1730 13584542-872b-42d8-b299-866967b5c3ef 31 15aece23-a215-4579-91b4-69eec72e18da32 15c3b339-88f7-4a86-ab16-e71c58dcb01e 33 1704f00f-79e6-43a7-961b-cedd3724d5fd34 1954cced-e748-45c4-9c26-9855b97fbc5e 35 1d17d234-e39d-4ed7-b46f-4417922a4e7c36 1de60575-b6e-4c3d-9e6a-2fa699f9f197 37 1e8df695-bd1b-45b3-b557-e7d599cf759738 1f18a87-af6f-41ef-9853-cdb8f32ebdea 39 20236825-b5df-46e7-89bf-62e1d640a89740 215dfd39-f493-4bc3-a027-8a97d72c61bf 41 21760ecb-8f62-40d2-8d85-0cee5725cb7242 21ab7b40-77c2-4ae6-8321-e00d3a086c73 43 227d2f97-562b-4ccb-ae47-a5ec9e142fbb44 23393935-50c7-4a86-aeea-2b78fd089c5c 45 236833a3-5704-47fc-888c-4f298f09f79946 2373b66a-092d-44cb-bfd7-82e86e7a3b4d 47 26150609-0da3-4a7d-8868-0faf9c5f01b48 26660ad1-6ebb-4f59-8cba-a8432dfe8d38 49 26a8440e-c166-4c50-aef4-bfb77314b46b50 276c624-87ea-4f08-ab93-f770e3790175 51 2888b4e6-5b47-4b57-8bf5-c7382789077452 28c3b7e-b194-4bc9-8353-d04c0f4d56d2 53 2ad9387a-65d8-4e33-ad5b-7580065a27ca54 2ae9ba84-3a0d-4d4c-8338-3a1478dc5fe3 55 2b9493d7-49b8-493a-a71b-56cd1f4d690856 2b94c692-6abb-48ae-ab0b-b3e8a19cb340 57 2c9fc0de-3e7-45e1-a5df-c86206ad78b558 2cd43775-7085-45d8-89fa-9e35c0a915cf 59 2fe4b718-3bd7-46ec-bdce-b184f565362460 3299584d-8f11-4457-bf4c-ce98f7600250 61 337d318b-a07-4f4f-b763-89d9a2d013f62 347ef137-7eeb-4c80-a3b-0951f26a8aff 63 35253b65-1c19-4304-8a4-6884b8218fc064 357ef137-7eeb-4c80-a3b-0951f26a8aff 65 36037439-2044-4b50-b9d1-875b5a33214366 3680a5e-6870-426a-a997-eba929a0d25c 67 368d9ba4-203c-40c1-9fa3-da2f1430ce6368 3720f614-37fd-4d04-8a6b-76f54f8c222d 69 37608790-6147-45d0-9f20-1137b35703d70 37887e8c-da15-4192-923c-08fa390a176d 71 386dbd0e-0241-4a0a-b6a2-6704fba26b1c72 39be0d19-634d-4475-8768-09c130f5425d 73 3a7c8185-25c1-4941-bd7b-96e823c9f21f74 3a93cae4-ad3e-403e-8c12-65303b271818 75 3a4e37-dc91-482e-99af-132a612d40f376 3c8f201a-009d-4bbe-8b65-a6f8b35b57f 77 3ce045a0-877b-42a-8d2c-b4a863336ab878 3e3fc409-bff3-4905-bf16-c968e3f807 79 3ef2b351-8a84-4f2-8724-d86eae9b842e80 3f05f3b9-29ba-4b6b-95a-2204697ffc06 81 3f28fe4f-5d9d-4994-a456-efd78cfae1a382 3f49d2c-f400-4e7d-90c-9b18e401c31 83 4127319a-8b79-4410-b58a-7a151e15f3d784 415ef462-bed3-493a-ac36-ca8c6d23bf1b 85 4172ea6e-6b77-4edb-a9c-c0014bd1603b86 4188d3a4-077d-46b7-9c86-23e1a036f6c1 87 42d25c08-fb87-4927-8b65-93631280a26f88 42e0a640-4f19-4b28-973d-729602b5a4a7 89 44e5668-ecd5-4366-a6ce-c1c9b8d4e93890 455d3c66-7dc6-4537-a39a-36d3e9119df7 91 47543840-672a-467d-80df-8f7c3b9788c992 47f7c0ce-a5fb-4100-a5e6-65cd0e7429e5 93 480bcfea-d68f-4a-a0a9-2589ef31938194 48c46dc7-fe04-4505-ade7-723cba1a6f6 95 48d05431-6cd5-4e76-82eb-12b60d823f7d96 4bcb1253-a636-4df4-8cb0-a35c04dfef31 33 Preprint Table 12: Complete task IDs for the 286-task OSWorld-Verified subset (continued). No.Task IDNo.Task ID 97 4c26e3f3-3a14-4d86-b44a-d3cedebbb48798 4d117223-a354-47fb-8b45-62ab1390a95f 99 4e60007a-f5be-4bfc-9723-c39affa0a6d3100 4ed5abd0-8b5d-47bd-839f-cacfa15ca37a 101 4f07fbe9-70de-4927-a4d5-b28bc12c52c102 510f64c8-9bcc-4be1-8d30-638705850618 103 51719eea-10bc-4246-a428-ac7c433d4b3104 51b11269-2ca8-4b2a-9163-f21758420e78 105 5203d847-2572-4150-912a-03f062254390106 535364ea-05bd-46ea-9937-9f55c68507e8 107 53ad5833-3455-407b-bbc6-45b4c79ab8fb108 550ce7e7-747b-495f-b122-acdc4d0b8e54 109 554785e9-4523-4e7a-b8e1-8016f565f56a110 57242fad-77ca-454f-b71b-f187181a9f23 111 57667013-ea97-417c-9dce-2713091e6e2a112 5812b315-e7bd-4265-b51f-863c02174c28 113 58565672-7bfe-48ab-b828-db349231de6b114 59155008-fe71-45ec-8a8f-dc35497b6a8 115 5990457f-2adb-467b-a4af-5c857c92d762116 5ac2891a-eacd-4954-b339-98abba077adb 117 5bc63fb9-276a-4439-a7c1-9dc76401737f118 5c1075ca-b34-46a3-a7a0-029bd7463e79 119 5ca86c6f-f317-49d8-b6a7-b527541caae8120 5ced85fc-fa1a-4217-95fd-0fb530545ce2 121 5cfb9197-e72b-454b-900e-c06b0c802b40122 5d901039-a89c-4bfb-967b-bf66f4df075e 123 5df7b33a-9f77-4101-823e-02f863e1c1ae124 5e2d93d8-8ad0-4435-b150-1692aacaa994 125 5ea617a3-0e86-4ba6-aab2-dac9a2e8d57126 6054afcb-5bab-4702-90a0-b259b5d3217c 127 62f7fd55-0687-4a43-b6e1-3eda16fc6252128 66399b0d-8fda-4618-95c4-bfc6191617e9 129 6766f2b8-8a72-417f-a9e5-56fcaa735837130 67890eb6-6ce5-4c00-9e3d-fb4972699b06 131 68a25bd4-59c7-4f4d-975e-da0c8509c848132 6a33f9b9-0a56-4844-9c3f-96ec3ffb3ba2 133 6ada715d-3aae-4a32-a6a7-429b2e43fb93134 6c4c23a1-42a4-43c-9db1-2f86f3738c 135 6e99a1ad-07d2-4b66-a1ce-ece6d99c20a5136 6f4073b8-d8ea-4ade-8a18-c5d1d5d5a9a 137 6f56bf42-85b8-4fbb-8e06-6c44960184ba138 6f81754e-285d-4ce0-b59e-af7edb02d108 139 70745df8-f2f5-42bd-8074-fbc10334fcc5140 72b810ef-4156-4d09-8f08-a0cf57e7cefe 141 72f83cdc-bf76-4531-9a1b-eb893a13f8a142 734d6579-c07d-47a8-9ae2-13339795476b 143 73c99fb9-f828-43ce-b87a-01dc07faa224144 74d5859f-ed66-4d3e-a0e-93d7a592ce41 145 7767eef2-56a3-4cea-8c9f-48c070c7d65b146 778efd0a-153f-4842-9214-f05fc176b877 147 77b8ab4d-994f-43ac-8930-8ca087d7c4b4148 788b3701-3ec9-4b67-b679-418bfa726c22 149 7a4e4bc8-922c-4c84-865c-25ba34136be1150 7a5a7856-f1b6-42a4-ade9-1ca81ca0f263 151 7ae48c60-f143-4119-b659-15b8f485eb9a152 7aeae0e2-70e-4705-821d-1bba5d5b2d 153 7b1e1f9-b85-49be-b01d-d6424be18cd0154 7b6c7e24-c58a-49fc-a5b-d57b80e5b4c3 155 7b7617bd-57c-468e-9c91-40c4ec2bcb3d156 7c4c09e-7a92-40d-8338-b2286535c4ed 157 7e287123-70ca-47b9-8521-47db09b69b14158 7e429b8d-a3f0-4ed0-9b58-08957d00b127 159 7f52cab9-535c-4835-ac8c-391e64dc930160 7f48d5b-2df2-49da-b500-a5150ffc7f18 161 81c425f5-78f3-4771-afd6-3d2973825947162 82279c77-8fc6-46f6-9622-3ba96f61b477 163 82bc8d6a-36eb-4d2d-8801-ef714fb1e55a164 82e3c869-49f6-4305-a7ce-f3e64a0618e7 165 841b50a-df53-47bd-a73a-22d3a9f73160166 8472fece-c7d-4241-8d65-9b3cd1a0b568 167 873cafdd-a581-47f6-8b33-b9696ddb7b05168 881deb30-9549-4583-a841-8270c65f2a17 169 88fe4b2d-3040-4c70-9a70-546a47764b48170 8979838c-54a5-4454-a2b8-3d135a1a5c8f 171 8ba5ae7a-5ae5-4eab-9fcc-5d4fe3abf89172 8f080098-ddb1-424c-b438-4e96e5e4786e 173 91190194-f406-4cd6-b3f9-c43fac942b22174 9195653c-f4a-453d-a95-787f6ccfaae9 175 930fdb3b-11a8-46fe-9bac-577332e2640e176 936321ce-5236-426a-9a20-e0e3c5dc536f 177 937087b6-f668-4ba6-9110-60682e33441178 9439a27b-18ae-42d8-9778-5f68f891805e 179 94d95f96-9699-4208-98ba-3c3119edf9c2180 9656a811-9b5b-4ddf-99c7-5117bcef0626 181 971cbb5b-3cbf-4f7-9e24-b5c84fcebfa6182 982d12a5-beab-424f-8d38-d2a48429e511 183 99146c54-4f37-4ab8-9327-5f3291665e1e184 9b7bc335-06b5-4cd3-9119-1a649c478509 185 9bc3c16-074a-45ac-9bdc-b2a362e1daf3186 9cf05d24-6bd9-4dae-8967-f67d88f5d38a 187 9d425400-e9b2-4424-9a4b-d4c7abac4140188 9ec204e4-f0a3-42f8-8458-b772a6797cab 189 9f3b592-209d-43bc-b47-d77d9df56504190 9f935cce-0a9f-435f-8007-817732bfc0a5 191 a01fbce3-2793-461f-ab86-43680ccbae25192 a097acff-6266-4291-9fbd-137af7ecd439 34 Preprint Table 12: Complete task IDs for the 286-task OSWorld-Verified subset (continued). No.Task IDNo.Task ID 193 a10b69e1-6034-4a2b-93e1-571d45194f75194 a1af9f1c-50d5-4bc3-a51e-4d9b425f638 195 a434992a-89df-4577-925c-0c58b747f0f4196 a462a795-fdc7-4b23-b689-e8b6df786b78 197 a4d98375-215b-4a4d-aee9-3d4370fccc41198 a503b07f-9119-456b-b75d-f5146737d24f 199 a53f80cd-4a90-4490-8310-097b011433f6200 a669ef01-ded5-4099-9ea9-25e99b569840 201 a728a36e-8bf1-4b6-9a03-ef039a5233f0202 a74b607e-6b5-4ea8-8a7c-5d97c7bbcd2a 203 a82b78b-7fde-4cb3-94a4-035baf10bcf0204 a96b564e-dbe9-42c3-9ccf-b4498073438a 205 a9f325a-8c05-4e4f-8341-9e4358565f4f206 a3a8974-2e85-438b-b29e-a64df44deb4b 207 a4b5023-aef6-4ed9-bdc9-705f59ab9ad6208 aad10cd7-9337-4b62-b704-a857848cedf2 209 abed40dc-063f-4598-8ba5-9fe749c0615d210 ac1b39f-e4d-4483-abce-c117e98942f0 211 acb0f96b-e27c-44d8-b55f-7cb76609dfcd212 aceb0368-56b8-4073-b70e-3dc9aee184e0 213 adf5e2c3-64c7-4644-b7b6-d2f0167927e7214 af23762e-2bfd-4a1d-aada-20fa8de9ce07 215 af630914-714e-4a24-a7b-f9af687d3b91216 b070486d-e161-459b-a2b-ef442d973b92 217 b148e375-fe0b-4bec-90e7-38632b0d73c2218 b337d106-053f-4d37-8da0-7f9c4043a66b 219 b3d4a89c-53f2-4d6b-8b6a-541fb5d205fa220 b4f95342-463e-4179-8c3f-193cd7241fb2 221 b5062e3e-641c-4e3a-907b-ac864d2e7652222 b6781586-6346-41cd-935a-a6b1487918fc 223 b7895e80-f4d1-4648-bee0-4eb45a6f1fa8224 b5e4c0d-f964-439c-97b6-bdb9747de3f4 225 b83cab4-e5c7-42c7-a67b-e46068032b86226 bba3381f-b5eb-4439-bd9e-80c22218d5a7 227 bc2b57f3-686d-4ec9-87ce-edf850b7e442228 bf4e9888-f10f-47af-8dba-76413038b73c 229 c1fa57f3-c3db-4596-8f09-020701085416230 c288e301-e626-4b98-a1ab-159dcb162af5 231 c59742c0-4323-4b9d-8a02-723c251deaa0232 c6bf789c-ba3a-4209-971d-b63abf0ab733 233 c7c1e4c3-9e92-4eba-a4b8-689953975ea4234 c82632a4-56b6-4db4-9d1-3820e3388e4 235 c867c42d-a52d-4a24-8ae3-f75d256b5618236 ce2b64a2-ddc1-4f91-8c7d-a88be7121aac 237 ce88f674-ab7a-43da-9201-468d38539e4a238 d1acdb87-b67-4f30-84a-990e56a09c92 239 d38192b0-17dc-4e1d-99c3-786d0117de77240 d52d6308-ec58-42b7-a2c9-de80e4837b2b 241 d53f5e-3b1a-431e-b2be-30ed2673079b242 d681960f-7bc3-4286-9913-a8812ba3261a 243 d9b7c649-c975-4f53-88f5-940b29c47247244 da46d875-6b82-4681-9284-653b0c7ae241 245 da52d699-e8d2-4dc5-9191-a2199e0b6a9b246 da922383-bfa4-4cd3-bbad-6bebab3d7742 247 dcbe20e8-647f-4f1d-8696-f1c5b570e3248 d60633f-2c72-42ba-8547-6f2c8cb0fdb0 249 d84e895-72fd-4023-a336-97689ded257c250 df67aebb-fb3a-44fd-b75b-51b6012df509 251 dfac9e8-9bc4-4cdc-b465-4a4bfcd2f397252 e0df059f-28a6-4169-924f-b9623e7184c 253 e135df7c-7687-4ac0-a5f0-76b74438b53e254 e1e75309-3ddb-4d09-92ec-de869c928143 255 e246f6d8-78d7-44ac-b668-fcf47946cb50256 e2b5e914-ffe1-44d2-8e92-58f8c5d92b2 257 e2d0213-26db-4349-abe5-d5667bfd725c258 e528b65e-1107-4b8c-8988-490e4fece599 259 ea98c5d7-3cf9-4f9b-8ad3-366b58e0fcae260 eabc805a-bfcf-4460-b250-ac92135819f6 261 eb03d19a-b88d-4de4-8a64-ca0ac66f426b262 eb303e01-261e-4972-8c07-c9b4e7a4922a 263 ec4e3f68-9ea4-4c18-a5c9-69f89d1178b3264 ec71221e-ac43-46f9-89b8-e7d80f7e1c5 265 ecb0df7a-4e8d-4a03-b162-053391d3afaf266 ecc2413d-8a48-416e-a3a2-d30106ca36cb 267 ed43c15f-00cb-4054-9c95-62c880865d68268 e9a3c83-f437-4879-8918-be5efbb9fac7 269 ef9d12bd-bcee-4ba0-a40e-918400f43ddf270 f0b971a1-6831-4b9b-a50e-22a6e47f45ba 271 f178a4a9-d090-4b56-bc4c-4b72a61a035d272 f201fbc3-44e6-46fc-bcaa-432f9815454c 273 f23acfd2-c485-4b7c-a1e7-d4303ddfe864274 f3977615-2b45-4ac5-8bba-80c17dbe2a37 275 f4aec372-4fb0-4df5-a52b-79e0e2a5d6ce276 f5d96daf-83a8-4c86-9686-bada31fc66ab 277 f723c744-e62c-4ae6-98d1-750d3cd7d79d278 f79439ad-3e8-4f99-a518-0eb60e5652b0 279 f7dfbef3-7697-431c-883a-db8583a4e4f9280 f8cfa149-d1c1-4215-8dac-4a0932bad3c2 281 f9584479-3d0d-4c79-affa-9ad7afdd8850282 f9be0997-4b7c-45c5-b05c-4612b44a6118 283 fba2c100-79e8-42df-ae74-b592418d54f4284 fbb548ca-c2a6-4601-9204-e39a2efc507b 285 fc6d8143-9452-4171-9459-7f515143419a286 fe41f596-a71b-4c2f-9b2f-9dcd40b568c3 35