菜单
Create or improve agent skills from concrete behavior and evidence. Use when building or revising SKILL.md, or testing whether a skill earns its context. Do not use for ordinary documentation or one-off prompts.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | craft-skill |
| description | Create or improve agent skills from concrete behavior and evidence. Use when building or revising SKILL.md, or testing whether a skill earns its context. Do not use for ordinary documentation or one-off prompts. |
A skill should be called for the right task, steer the agent to genuine completion, and make every token earn its place.
Read the request, conversation, existing skill, repository conventions, and target platform's authoritative skill specification.
Define the useful behavior with realistic examples: what should call the skill, what should not, what the agent should do differently, and what observable evidence means it finished. Ask only about user-owned choices that materially change those answers. When revising a skill, preserve its name and inspect observed failures before editing.
SKILL.md. Put branch-specific knowledge behind a direct pointer that states when to load it.Call is complete when every intended branch and close non-trigger has one clear owner, and the metadata passes the platform's validator when one exists.
Use the lightest reliable medium:
Match freedom to fragility. Use principles where several approaches work; tighten procedure or code only where variance causes failure.
Define completion from the requested outcome, not the procedure. Exhaustive means every decision-relevant item is accounted for, not every discoverable source or intermediate step is performed. End each step with a completion criterion the agent can check, with honest terminal states for missing evidence, unavailable input, and blocked actions. End with an overall postcondition that inspects the resulting artifact or state; completed steps are not final evidence.
When evidence could change the decision, read references/evaluation.md before evaluating. When an existing skill misbehaves and the cause is unclear, read references/diagnosis.md.
Run the platform validator and relevant script tests. Confirm every resource has a condition-bearing pointer. Report the evidence and what remains unverified.
Stop when the concrete examples work and another deletion would damage Call, Steer, or the evidence needed to judge them.
Use this skill whenever code-review feedback already exists and may affect a code decision or change. This includes PR comments, findings returned by invoke or any external model or review tool, and delegated-review results—even when the user simply says to fix, address, apply, or resolve them. Do not use to perform a new review, diagnose an unrelated bug, implement a new requirement, or merely summarize or explain feedback without deciding about code.
Interview the user to resolve material branches of a decision tree and converge on a shared plan before acting. Use when multiple reasonable choices would materially change scope, cost, reversibility, or outcome, or when the user asks to brainstorm, shape, or grill an idea. Do not use when facts can resolve the ambiguity or a reversible assumption would not materially affect the result.
Draft or review minimal, testable goals. Use only when the user explicitly asks to write, revise, or review a goal.
Diagnose and fix bugs, failures, and performance regressions. Read this skill BEFORE touching any code whenever the user reports a bug, broken behavior, a crash, wrong output, or a regression — even when the cause looks obvious and the user only asked for a fix. Do not use for a simple explanation of a known error when no diagnosis is requested.
Research with parallel agents, independent evidence angles, and cross-validation. Use when the user explicitly asks for parallel or delegated research, or when a high-cost question materially benefits from independent evidence.
Teach with adaptive guidance, retrieval practice, and Socratic questioning. Use when the user wants to study a topic, build durable understanding, practice, or be coached. Do not use for one-off factual answers, quick explanations, or ordinary task guidance.