Extract behavior-changing ways of working, not a project summary. Project outcome is context, not the criterion: a failed project may yield a strong procedure, while a successful one may yield none.
Workflow
- Inspect the smallest useful evidence set: instructions, source, tests, scripts, commands, logs, errors, reviews, history, and outputs. State what was unavailable; do not imply it was inspected.
- Identify repeatable actions: safe tool/editing sequences, diagnosis methods, validation loops, decision or stop rules, and reproducible environment patterns. Exclude lore, wishes, TODOs, ordinary setup, motivation, and one-off facts.
- Recommend a skill only when the candidate has:
- a distinct trigger and repeatable procedure;
- plausible transfer to several future tasks after project names are removed;
- behavior-changing guidance beyond “be careful”;
- a concrete success check tied to an observed or credible failure;
- a better home in a skill than in a README, comment, issue, or postmortem;
- a focused scope.
- Search the current skill collection before proposing a directory. Update when trigger, workflow, and validation substantially match; merge project-named variants; split candidates whose triggers, tools, procedures, or checks differ.
- Generalize names and paths while preserving causal rules, ordering, tool constraints, failure signals, and validation. Keep tool commands only for tool-specific skills. Put conditional detail and long examples in directly linked
references/files. - Draft and review at most three strong recommendations by default. Use output-templates.md for candidate, draft, and report formats, then use failure-modes.md to reject or revise anti-patterns. If reachable, review current official Agent Skills authoring guidance; otherwise report that limitation rather than blocking.
- After drafting, use empirical-tuning-gate.md to decide whether fresh-executor testing is worth its cost. Structural self-review is not empirical validation.
Prefer zero strong candidates to several weak ones. Do not create skills merely to salvage a project or satisfy a requested quantity.
Final State
Report create/update/merge/reject decisions, intentionally excluded material, and remaining validation. Label each result honestly:
- Draft — extracted but untested.
- Structurally reviewed — checked for scope, clarity, links, and validation design.
- Empirically tuned — exercised by fresh executors and revised from observed failures.
- Rejected — not worth creating or merging.