writing-skills
Create, refactor, or debug agent skills and their loading. Use writing-agents-md for AGENTS.md/CLAUDE.md; excludes general docs and READMEs.
- License 许可
- MIT
- 所属仓库
- compozy/compozy
- 最近更新
- 2026-11-15
- Stars
- ★ 2.9k
- Forks
- 187
01 前言 Front Matter
- name
- writing-skills
- metadata
- { "author": "Pedro Nauck", "github": "https://github.com/pedronauck", "credits": "Doctrine adapted from writing-great-skills (mattpocock/skills)", "repository": "https://github.com/pedronauck/skills" }
- description
- Create, refactor, or debug agent skills and their loading. Use writing-agents-md for AGENTS.md/CLAUDE.md; excludes general docs and READMEs.
02 资源文件 6
- 资产文件assets/SKILL.template.md663 B
- 参考文档references/authoring-procedure.md1.3 KB
- 参考文档references/checklist.md1.5 KB
- 参考文档references/glossary.md3.1 KB
- 参考文档references/loading-diagnosis.md1.3 KB
- 脚本scripts/validate-metadata.py2.1 KB
03 SKILL.md 正文
Writing Skills
A skill supplies task-specific knowledge or a reliable operation the model would otherwise have to rediscover. Optimize for correct outcomes, useful evidence, and low overhead. Identical process on every run is useful only where ordering is part of the contract.
Choose the Scope
- Follow the user's requested outcome and existing authorization. A skill does not add approval steps or override repository policy.
- Keep a rule only when it changes a relevant decision or prevents an evidenced failure. Delete generic advice, duplicated rules, obsolete model assumptions, and rhetorical pressure.
- Keep descriptions short: put the concrete task, trigger, and main scope boundary in the first sentence so truncation preserves when to use the skill. Add exclusions only to resolve likely overlap. Avoid firing on every code edit, every test, or every completion when the procedure is only needed for a narrower problem.
- Distinguish shared guidance from repository-specific policy. Edit shared sources first; preserve local variants when synchronizing. Check whole-directory differences before an installer overwrites a skill.
Create or Revise
- Identify the outcome, owning files, known failure, and existing instructions. Reuse the current conversation's evidence; recollect only missing or stale material.
- Keep the minimum common contract in
SKILL.md; link branch-specific references with a condition and the section needed. Read a complete schema or tightly coupled procedure when its whole contract is necessary, not every reference by default. - State required order only for dependencies, mutations, authorization, or cleanup. Let independent searches batch and let the agent choose routine implementation details.
- Define completion by the requested deliverable and its applicable checks. Reuse evidence for unchanged inputs. Do not add a mandatory report, subagent round, test suite, or confirmation merely because the skill is installed.
- Check changed triggers, links, metadata, and any affected helper behavior. For a small prose edit, summarize this review; for a new skill or broad rewrite, use
references/checklist.md. Fix material failures at their source.
For a new package, read references/authoring-procedure.md; assets/SKILL.template.md is an optional skeleton. Resolve bundled helpers from the skill's actual directory. The read-only metadata helper is python3 <writing-skills-dir>/scripts/validate-metadata.py --name "[name]" --description "[description]".
Diagnose and Measure
If a necessary reference was missed, use references/loading-diagnosis.md to distinguish a bad trigger, inaccessible path, conflicting rule, stale content, and model error. Strengthen the relevant pointer or fix the helper; extra emphasis is not the default repair.
Evaluate behavior-changing guidance on representative tasks under the same model/harness when practical: correctness, rework, unnecessary questions, tool calls, and elapsed time. Include successful runs as controls. Count reference reads and skill invocations as costs or diagnostics, never success by themselves. State when only static review was performed; do not claim measured improvements without comparative runs.
Terminology, when needed: references/glossary.md.
04 同仓库更多 Skills
security-review
Security code review for vulnerabilities. Use when asked to "security review", "find vulnerabilities", "check for security issues", "audit security", "OWASP review", or review code for injection, XSS, authentication, authorization, cryptography issues. Provides systematic review with confidence-based reporting.
lesson-learned
Extract evidence-backed engineering lessons from recent changes or incidents. Use for requested reflection, lesson updates, or retiring an obsolete workflow rule.
compozy
Operate CompozyOS. Use when working with CompozyOS sessions, agents, native tools, skills, tasks, Loops, Goals, Terminal, desktops and windows, automation, extensions, or configuration. Don't use for unrelated projects.
remotion-upgrade
Upgrade Remotion, and related packages
remotion-studio
Preview a Remotion video
remotion-saas
Build an app with Remotion