Decomposing agent traces into workflow, semantics, and attachments beats prompted summaries by 10.5%

Notes on Workflow-to-Skill: Skill Creation via Routing-Workflow-Semantics-Attachments Decomposition (arXiv:2606.06893) — Yuyang Zhang, Xinyuan Han, Xudong Jiang, Runzhe Wang · June 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

Most Skill-authoring pipelines still turn a pile of agent traces into a single block of prose. RWSA (Routing–Workflow–Semantics–Attachments) is a representation built to stop that: it splits a skill into a routing header — when the skill applies — plus three separate runtime layers: workflow (control flow, branches, verification checkpoints), semantics (each step's objective and decision rule), and attachments (tools, validation, rollback, state management). The accompanying W2S pipeline builds this bottom-up from raw evidence. It segments traces, drafts a local skill per segment, then aligns shared structure across traces of the same task — reconciling branch cases instead of averaging them away, and compressing duplication while keeping every claim tagged back to the evidence and confidence level that produced it. Before it ships, a feedback pass hunts for holes: missing branches or criteria, steps that are vague or out of order. It also flags contradictions between the routing, workflow, and semantics layers.

Tested on WSASkill, a new 70-skill benchmark spanning all eight combinations of workflow, semantics, and attachment presence, with 10 real interaction traces backing every execution path. Fidelity is measured by behavioral replay — whether an agent handed the generated skill actually reproduces the reference behavior, not whether the resulting text reads well. W2S beats Anthropic Skill Creator (a structured-interview-and-drafting baseline) on seven of eight skill types, averaging 0.503 versus 0.455 fidelity — a 10.5% relative gain. The advantage is largest where a skill's value is almost pure decision logic, and it flips to a loss on the one type built from workflow and attachments without explicit semantics: there, prompting for a plain summary beats forcing structure onto a skill with little decision logic to extract. Even the strongest category tops out around three-quarters fidelity, and the hardest skill type — workflow only, no semantics or attachments — barely clears a quarter for W2S (0.276) and falls short of it for the baseline (0.227).

Key numbers

Benchmark scale	70 skills across 8 structural types
Replay fidelity gain	+10.5% relative vs. prompted baseline
Evidence depth	10 traces collected per execution path
Where structure backfires	-0.070 vs. baseline on workflow+attachments-only skills
Fidelity ceiling	0.503 avg (of 1.0) even for the winning method

What this means for skill authors

For skill authors, the practical takeaway is not subtle: don't hand a raw trace log to an LLM and ask for a README-style skill if the behavior you're capturing depends on branching logic, verification steps, or rollback conditions. A prompted summary is measurably worse at reproducing that behavior than a method that keeps workflow, semantics, and attachments as separate, checkable objects. That's the same distinction the skill-authoring community keeps re-learning the hard way: a skill is a runtime spec an agent executes, not a document a human reads.

The one reversal in the results is worth sitting with. When a skill is mostly "here are the tools and constraints" with little decision logic underneath, decomposition doesn't help and can actively hurt — forcing structure onto something that's already thin on logic adds overhead instead of signal. That's a caution against over-engineering simple skills, and it lines up with why a vetted catalog benefits from checking a skill's behavior in situ rather than assuming more structure always means more reliability.

It's also a reminder that automated skill construction is progress on a real cost problem, not a finished one. The winning method here still caps out well short of perfect replay fidelity across most skill types, which means anything produced this way still needs behavioral verification before it earns a place in an agent's tool belt — text-only review of a skill's prose won't catch a workflow that only fails on its third branch.

Skills related to this research

nodemailer nodemailer enables Node.js developers to dispatch transactional emails programmatically with straightforward configuration and robust delivery handling. Whether you're sending password resets, order confirmations, or notifications, this skill integrates email functionality directly into your application workflow.★ 84 axiom-run-tests axiom-run-tests equips AI coding assistants with specialized capabilities for executing and debugging Apple platform tests. It surfaces actionable diagnostics when XCUITests fail, helping you identify root causes and fix issues faster without manual investigation overhead.★ 1,095 playwright-scanner playwright-scanner automates accessibility testing by running behavioral checks against live web pages through Playwright. It identifies WCAG compliance gaps and accessibility barriers, helping teams catch usability issues before users encounter them. Powered by Gemini for intelligent analysis.★ 370 seedance-recipes Seedance-recipes provides production-ready recipe patterns for video content across genres: product, lifestyle, drama, music video, landscape, commercial, animation, and more. Each recipe preserves core creative constraints while inviting customization of subject, camera, lighting, and sound. Use recipes as proven starting shapes, not rigid templates.★ 5,445

Related notes

References

Zhang, Han, Jiang & Wang, "Workflow-to-Skill: Skill Creation via Routing-Workflow-Semantics-Attachments Decomposition" (arXiv:2606.06893, 2026)
Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models" (2022)
Shinn et al., "Reflexion: Language Agents with Verbal Reinforcement Learning" (2023)
Xie et al., "OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments" (2024)
Wang et al., "Agent Workflow Memory" (2024)