AgentSkillsCN

verification-before-completion

在未获得最新、确凿的证据之前，严禁做出任何关于成功或完成的宣称。每一项声明都必须以当前会话中实际执行的命令输出为依据。

SKILL.md
--- frontmatter
name: verification-before-completion
description: Verification protocol that prohibits any success or completion claims without fresh, concrete evidence. Every claim must be backed by actual command output from the current session.
autoApply: true

Verification Before Completion

Core Mandate

NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE.

Never say "done," "complete," "passing," or "working" without running the verification commands in this session and showing the output. Memory of a previous run does not count. Assumptions do not count. Only fresh evidence counts.

The Five-Step Gate

Every completion claim must pass through all five steps in order:

1. Identify

Determine exactly what needs to be verified:

  • •Which tests must pass?
  • •Which build commands must succeed?
  • •Which linters or formatters must report clean?
  • •Which endpoints must respond correctly?
  • •Which acceptance criteria must be met?

Write the list down before running anything.

2. Run

Execute every verification command. Do not skip any.

bash
# Example verification sequence
pytest tests/ -v              # unit and integration tests
python -m mypy src/           # type checking
python -m flake8 src/         # linting
python -m build               # build succeeds

Run them now, in this session. Do not rely on earlier runs.

3. Read

Read the full output of every command. Do not scan for "PASSED" and move on.

  • •Check for warnings that indicate hidden problems.
  • •Check for skipped tests that should have run.
  • •Check for deprecation notices that may cause future failures.
  • •Check that the number of tests matches expectations.

4. Verify

Confirm that every item on the identification list from Step 1 is satisfied:

  • •All specified tests pass.
  • •Build completes without errors.
  • •Linters report clean (or only known/accepted warnings).
  • •All acceptance criteria are met.

If any item fails, do not proceed to Step 5. Fix the issue and restart from Step 2.

5. Claim

Only now may you claim completion. When you do:

  • •State what was verified, not just "it works."
  • •Include the evidence -- paste command output or key excerpts.
  • •Be specific -- "All 47 tests pass" not "tests pass."

Rules

Evidence Before Claims, Always

The order is always: run command, read output, then make claim. Never the reverse.

Never Claim Tests Pass Without Running Them

"The tests should pass" is not verification. Run them.

Never Assume Linter Success Means Build Success

They check different things. Run both.

Never Use Hedging Language

These phrases are banned in completion claims:

  • •"should work"
  • •"probably passes"
  • •"I believe this is correct"
  • •"this looks good"
  • •"I think we're done"

Replace them with evidence-backed statements:

  • •"All 47 tests pass (output above)."
  • •"Build completed successfully with exit code 0."
  • •"Linter reports 0 errors, 0 warnings."

Applies to ALL Success Claims

This protocol applies every time you claim something is:

  • •Complete
  • •Fixed
  • •Working
  • •Passing
  • •Ready for review
  • •Done

No matter how small the change, no matter how confident you are, run the verification.

No Shortcuts

  • •"I just ran it a minute ago" -- Run it again.
  • •"It's a trivial change" -- Verify it.
  • •"The test already existed" -- Run it to confirm it still passes.
  • •"Nothing else could have broken" -- Prove it.

Verification Report Template

When reporting completion, use this structure:

code
## Verification Report

### Commands Run
1. `pytest tests/ -v` -- 47 passed, 0 failed, 0 skipped
2. `flake8 src/` -- 0 errors, 0 warnings
3. `python -m build` -- exit code 0

### Evidence
<paste key output sections>

### Conclusion
All verification checks passed. Implementation is complete.

Adapted from obra/superpowers