AgentSkillsCN

verifying-completion

强制对刚刚完成的工作进行细致的自我审查。逐条梳理每一处变更,生成验证待办事项,随后借助结构化的不确定性分析,精准查找潜在漏洞。在完成实施工作后、正式宣告任务“已完成”之前,务必使用此技能。

SKILL.md
--- frontmatter
name: verifying-completion
description: Forces meticulous self-review of work just completed. Itemizes every
  change, creates verification todos, then hunts for gaps using structured uncertainty
  analysis. Use after finishing implementation work, before claiming done.
required_todos:
- itemize-everything
- verification-todos
- gap-hunt
- verdict
<role> WHO: Completion auditor ATTITUDE: "Done" without verification is a lie. </role> <purpose> Your job is to prove work is actually complete. Itemize ruthlessly, verify each piece, then hunt what you missed. </purpose> <workflow> ## Phase 1: Itemize Everything

BLOCKING GATE: You just completed implementation work in this session.

Review EVERYTHING you did. Not summaries. Actual items:

CategoryWhat to List
Files createdEvery new file, with purpose
Files modifiedEvery edit, what changed and why
Code addedFunctions, classes, methods - each one
Code removedWhat was deleted and why
Config changesSettings, env vars, dependencies
TestsTest files, test cases
Commands runBuild, test, lint results

Be paranoid. If you touched it, list it. If you thought about touching it but didn't, note that too.

EXIT CRITERIA: Written inventory of every discrete piece of work.


Phase 2: Verification Todos

BLOCKING GATE: Phase 1 inventory complete.

Convert each item to a verification todo. Use TodoWrite:

python
TodoWrite(todos=[
    {"content": "Verify: [specific item] is complete", "status": "pending", "activeForm": "Verifying [item]"},
    # One todo per discrete piece of work
])

Verification means:

  • Code compiles/runs without error
  • Tests pass (if applicable)
  • No TODO/FIXME left behind
  • No placeholder implementations
  • Imports resolved
  • Types match (if typed)

Work through each todo. Mark complete only with evidence.

EXIT CRITERIA: All verification todos complete OR blocked items identified.


Phase 3: Gap Hunt

BLOCKING GATE: Phase 2 verification complete.

Now challenge yourself using structured uncertainty:

UNKNOWNS: (list before analyzing)

  • What did I assume was obvious?
  • What edge cases didn't I test?
  • What error paths didn't I handle?
  • What dependencies did I assume exist?
  • What did the user ask for that I might have interpreted differently?

ALTERNATIVES: (ways this could have been done differently)

  • What if I misunderstood the scope?
  • What if there's a simpler solution I missed?
  • What if I over-engineered something?

GAPS FOUND: (be honest) For each unknown, check:

  • Resolved - evidence: [what proves it]
  • Still unknown - impact: [what could break]
  • Gap found - action: [what to fix]

EXIT CRITERIA: All unknowns addressed. Gaps either fixed or explicitly flagged.


Phase 4: Verdict

State clearly:

COMPLETE: All items verified, no gaps found.

OR

INCOMPLETE: [List specific items that need attention]

No hedging. No "should be fine." Evidence or gap. </workflow>

<rules> - Every item gets a todo. No batching "miscellaneous changes." - "Looks complete" is not verification. Run it, test it, prove it. - Finding gaps is success, not failure. Missing them is failure. - If you can't verify something, say so. Don't assume. - User asked for X - did you deliver X? Not your interpretation of X. </rules>