skills/verification-before-completion/SKILL.md
Use when claiming a task is complete, before commits, or before PR creation.
npx skillsauth add seokan-jeong/team-shinchan team-shinchan:verification-before-completionInstall this skill globally with one command. Works with Claude Code, Cursor, and Windsurf.
3 of 9 scanners reported clean
Some scanners were skipped, did not run, or reported a non-clean status. Review each row below.
Never claim a task is complete without running a verification command and reading its output. "I believe it works" is not evidence.
You MUST verify at each of these triggers:
Every completion claim must include:
If you cannot provide all three, you have NOT verified. Go back and run the command.
If your completion report contains ANY of these phrases, you have NOT verified -- STOP and actually run a command:
| Red Flag Phrase | What To Do Instead | |-----------------|--------------------| | "should work" | Run the test and show it passes | | "probably fine" | Run the linter/build and show zero errors | | "seems to" | Read the actual output and report what it says | | "looks correct" | Run a validator or test that proves correctness | | "I believe" | Replace belief with evidence -- run a command | | "likely" | Remove doubt -- execute and observe | | "I think" | Verify instead of thinking | | "no issues expected" | Prove it -- run the check |
Use this format in your completion report:
### Verification
- **Command**: `node tests/validate/index.js`
- **Expected**: All validators pass, 0 errors
- **Actual**: 23 validators passed, 0 errors, 0 warnings
- **Verdict**: PASS
If the verdict is FAIL, fix the issue before claiming completion. Do not report a task as done with a FAIL verdict.
At the before-PR and before-completion checkpoints, an APPROVED fierce-review artifact satisfies the code-review evidence requirement (it is a schema-validated, adversarially-verified rubric pass — stronger than a single read-through). If .shinchan-docs/reviews/REVIEW-{NNN}.json exists for the current scope with verdict.verdict == "APPROVED" and an empty must_fix, cite it:
/team-shinchan:fierce-review <scope>.shinchan-docs/reviews/REVIEW-{NNN}.jsonA REJECTED artifact, a non-empty must_fix, or a stale one (scope no longer matches the current diff) is NOT evidence — re-run the review. This does not replace the test/build evidence above; it adds the review dimension.
development
Deterministic adversarial code review for high-stakes scope — independent per-dimension review, a non-skippable per-finding refutation, a completeness critic, and a schema-validated rubric judge. Opt-in main-loop Workflow tier.
data-ai
Deterministic loop-until-done for high-stakes long-running tasks — a worker/verifier loop the script bounds by iteration cap, token budget, and stagnation, closed by an Action-Kamen gate. Opt-in main-loop Workflow tier.
testing
Deterministic adversarial debate for high-stakes or irreversible decisions — mandatory refutation plus a scored judge panel. Opt-in main-loop Workflow tier.
development
Deterministic competitive code tournament — N builders independently solve one task and return patches, an Action-Kamen judge scores them head-to-head, the winner is picked by score and applied. Opt-in main-loop Workflow tier.