界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
开发与工程 · SKILL.md

update-skill-content

update-skill-content

【开发与工程】Update the content of an existing SKILL.md to improve accuracy, completeness, and clarity. Covers version bumping, procedure refinement, pitfall expansion, and related skills synchronization. Use when a skill's procedures reference outdated tools or APIs, the Common Pitfalls section is thin, Related Skills has broken cross-references, or after receiv…

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: update-skill-content locale: caveman-ultra source_locale: en source_commit: 82c77053 fence_basis_commit: 82c77053 translator: "Julius Brussee homage — caveman" translation_date: "2026-05-03" description: > Update the content of an existing SKILL.md to improve accuracy, completeness, and clarity. Covers version bumping, procedure refinement, pitfall expansion, and related skills synchronization. Use when a skill's procedures reference outdated tools or APIs, the Common Pitfalls section is thin, Related Skills has broken cross-references, or after receiving feedback that a skill's procedures are unclear or incomplete. license: MIT allowed-tools: Read Write Edit Grep Glob metadata: author: Philipp Thoss version: "1.1" domain: review complexity: intermediate language: multi tags: review, skills, content, update, maintenance, quality

Update Skill Content

Improve existing SKILL.md → refine procedure, expand pitfalls w/ real failure modes, sync Related Skills, bump ver. Use after format passes but content has gaps, stale refs, incomplete procedures.

Use When

  • Procedure refs outdated tools|APIs|vers
  • Pitfalls thin (< 3) or missing real failure modes
  • Related Skills broken refs|missing relevant
  • Procedure lacks concrete code|vague instructions
  • New skill added → existing should cross-ref
  • Feedback → procedures unclear|incomplete

In

  • Required: Path to SKILL.md
  • Optional: Specific section(s) ("procedure", "pitfalls", "related-skills")
  • Optional: Update source (changelog, issue, feedback)
  • Optional: Bump ver (default: yes, minor)

Do

Step 1: Read + Assess

Read entire SKILL.md, eval each section.

Per-section criteria:

  • When to Use: Triggers concrete + actionable? (3-5 items)
  • Inputs: Types, defaults, required|optional clear?
  • Procedure: Each step has concrete code, Expected, On failure?
  • Validation: Checklist objectively testable? (5+)
  • Pitfalls: Specific w/ symptoms + fixes? (3-6)
  • Related Skills: Refs exist? Obvious related missing?

Got: Clear picture of which sections need work, specific gaps ID'd.

If err: Can't read (path err) → verify path. Broken YAML frontmatter → fix first via review-skill-format before content updates.

Step 2: Stale Refs

Scan procedure for ver-specific refs, tool names, URLs, API patterns.

Staleness indicators:

  • Specific vers (v1.24, R 4.3.0, Node 18)
  • URLs maybe moved|expired
  • CLI flags|cmd syntax changed
  • Pkg names renamed|deprecated
  • Config formats evolved
# Check for version-specific references
grep -nE '[vV][0-9]+\.[0-9]+' skills/<skill-name>/SKILL.md

# Check for URLs
grep -nE 'https?://' skills/<skill-name>/SKILL.md

Got: List of stale refs w/ line nums. Each verified current or flagged.

If err: Too many to check manually → prioritize: procedure code blocks first (most likely runtime fail), then Pitfalls (old workarounds), then info text.

Step 3: Update Procedure

Per step needing improvement:

  1. Verify code blocks exec correctly|reflect best practices
  2. Add missing ctx → explain why
  3. Concrete cmds: real paths, real flags, real out
  4. Update Expected → match current tool behavior
  5. Update On failure → current err msgs + fixes

When updating code, preserve orig structure:

  • Step numbering consistent
  • ### Step N: Title format
  • No reorder unless orig wrong

Got: All procedure steps current + executable code. Expected/On failure reflect actual current behavior.

If err: Unsure code still correct → add note <!-- TODO: Verify this command against current version -->. No remove working code → untested replace.

Step 4: Expand Pitfalls

Review Pitfalls, expand if gaps.

Quality:

  • Each: bold name + specific description
  • Description: symptom (what wrong) + fix (avoid|recover)
  • Real failure modes, not hypothetical
  • 3-6 target

Sources for new:

  • Procedure steps w/ complex On failure (likely pitfalls)
  • Related skills warning about same tools|patterns
  • Common user-reported issues

Got: 3-6 pitfalls, each w/ specific symptom + fix. No generic "be careful"|"test thoroughly".

If err: Only 1-2 ID'd → OK for basic. Intermediate|advanced w/ < 3 → author hasn't fully explored failure modes → flag for future.

Step 5: Sync Related Skills

Verify cross-refs valid + add missing.

  1. Per ref'd skill → verify exists:
    # Check if referenced skill exists
    test -d skills/referenced-skill-name && echo "EXISTS" || echo "NOT FOUND"
    
  2. Skills referencing this (should cross-link):
    # Find skills that reference this skill
    grep -rl "skill-name" skills/*/SKILL.md
    
  3. Check obvious related → domain + tags
  4. Format: - \skill-id` — one-line description of relationship`

Got: All ref'd skills exist on disk. Bidirectional refs in place. No orphaned links.

If err: Ref'd doesn't exist → remove ref or note as planned future. Many skills ref this but missing from Related → add 2-3 most relevant.

Step 6: Bump Ver

Update metadata.version per semver:

  • Patch (1.0 → 1.1): Typos, minor clarifications, URL updates
  • Minor (1.0 → 2.0): New procedure steps, significant additions, structural changes
  • Note: Skills use simplified two-part (major.minor)

Update date fields if present.

Got: Ver bumped appropriately. Change magnitude matches scope.

If err: Current ver unparseable → set "1.1" + comment noting history gap.

Check

  • All procedure steps current + executable code|concrete instructions
  • No stale ver refs, URLs, deprecated tool names
  • Every step has Expected: + On failure: blocks
  • Pitfalls 3-6 specific w/ symptoms + fixes
  • All Related Skills refs → existing skills
  • Bidirectional refs for closely related
  • Ver bumped appropriately
  • Line count < 500
  • SKILL.md still passes review-skill-format after changes

Traps

  • Update code w/o testing: Changing cmd w/o verify works → worse than leaving old. Uncertain → add verify comment, not untested replace.
  • Over-expand pitfalls: 10+ dilutes section. Keep 3-6 most impactful → edge cases → references/.
  • Break refs during updates: Renaming|domain change → grep entire library. grep -rl "old-name" skills/ finds all.
  • Forget bump: Every update, no matter small, bump ver. Lets consumers detect changes.
  • Scope creep → refactor: Content updates improve what skill says. Restructuring|extracting → switch to refactor-skill-structure.

→

  • review-skill-format — Run format validation before content updates
  • refactor-skill-structure — Content updates push > 500 lines → refactor structure for room
  • evolve-skill — Deeper changes beyond content updates (advanced variant)
  • create-skill — Reference canonical format spec for new sections|steps
  • repair-broken-references — Bulk cross-ref repair across library
我的收藏