Verify work
Prove a change works with real checks before calling it done.
$ npx skills add tonygorb/skills --skill verify-workPick your agents from a list.
$ npx skills add tonygorb/skills --skill verify-work -g -a claude-code -yInstalls into ~/.claude/skills/.
Open CustomizeSkills and upload it.
$ npx skills add tonygorb/skills --skill verify-work -g -a codex -yInstalls into ~/.agents/skills/.
$ npx skills add tonygorb/skills --skill verify-work -g -a cursor -yInstalls into ~/.agents/skills/.
$ npx skills add tonygorb/skills --skill verify-work -g -a github-copilot -yInstalls into ~/.agents/skills/.
$ npx skills add tonygorb/skills --skill verify-work -g -a gemini-cli -yInstalls into ~/.agents/skills/.
$ npx skills add tonygorb/skills --skill verify-work -g -a opencode -yInstalls into ~/.agents/skills/.
$ npx skills add tonygorb/skills --skill verify-work -g -a windsurf -yInstalls into ~/.codeium/windsurf/skills/.
Your agent uses it when you say
- “check it works”
- “run the checks”
What it does
Evidence before claims. Never report "done" from memory: run the commands, read the output. Validation scales with what the change can break, never with how many lines it is.
How it works
- 1Tier 1: always (seconds, terminal only)
- 2Tier 2: only if the UI changed
- 3Tier 3: only when asked
- 4Close out
Verify work
Evidence before claims. Never report "done" from memory: run the commands, read the output. Validation scales with what the change can break, never with how many lines it is.
Tier 1: always (seconds, terminal only)
Run the project's own checks. Find them in package.json, a Makefile or the README; typically:
- lint
- typecheck
- unit tests
All must pass. Report real numbers ("212/212 tests"), not "tests pass". A failure means fix, re-run, and only then report.
Tier 2: only if the UI changed
Skip it for pure logic, data or migration changes.
- Start the dev server if it isn't running.
- Open only the screens you changed. No crawling.
- Take two screenshots: desktop (1440 wide) and phone (390 wide).
- Look at them: overflow, clipped text, broken alignment, missing states, focus you can't see.
- If you need to log in and don't have test credentials, ask. Never guess or hard-code them.
Tier 3: only when asked
End-to-end suites and anything slow, costly or touching live data. If you think it's needed, say so and let the user decide.
Close out
- Report each tier: what ran and the real result, plus the screenshots if you took them.
- List the files you changed. Don't commit or push unless the user asked you to.
Red flags: you're about to skip verification
- "It's a tiny change." Tiny changes break builds. Tier 1 takes seconds.
- "Tests passed earlier." Earlier isn't after your last edit. Run them again.
- "The UI change is obviously fine." Look at it in the browser before you say so.
Credits
The "evidence before claims" principle and the red-flags format come from superpowers by Jesse Vincent (MIT).