| name | verify-work |
| description | Choose honest proof for a work claim and reject proof that could pass while the claim is false. Use before claiming non-trivial work done or when tests, runtime evidence, screenshots, performance, security, CI, static checks, or repo inspection must prove behavior, cleanup, migration, or replacement. Use this skill for proof selection only; specialist skills own proof mechanics. |
Verify Work
Owns proof selection and proof honesty. Specialist skills and plugins own how
to obtain platform, test, security, performance, screenshot, browser, or mobile
evidence.
Rule
Do not claim non-trivial work is done without proof that would fail if the
claim were false.
Start from the claim, not from convenient commands. Pick the smallest proof
surface that honestly covers the claim and material risks.
For non-trivial work, the claim must include the expected finished repo state
or an explicit accepted reduction. Do not prove only the latest edit, failing
test, or PR comment when the user objective requires a broader state to be
true.
Use unit tests, integration tests, e2e/browser/mobile flows, runtime evidence,
screenshots, performance traces, security scans, CI logs, static checks, or
repo inspection as appropriate. Read specialist skills in tandem for mechanics.
When proof adds, edits, deletes, reviews, or relies on persistent tests, use
../testing-best-practices/SKILL.md.
Proof is invalid when it could pass while the claimed behavior, cleanup,
migration, replacement, or integration is absent.
Cleanup and migration need positive replacement/reachability proof plus
supporting absence checks. Text-not-present, class-gone, or grep-only evidence
is not enough unless the claim is only that exact identifier was removed.
For migration, runtime, integration, or workflow failures, prefer rerunning the
real failed path or a faithful smoke/integration proof before adding mocked
persistent tests. A fake test around emitted commands, DDL, mocks, or call
choreography is not proof of the runtime boundary.
Verdict
Report claim, proof surface, commands/actions, material artifacts, verdict
pass | reject | blocked, limits, and remaining unproved boundaries.
Return blocked when the right proof surface is unavailable, unsafe, too
expensive for the current turn, or requires missing user/project input. Do not
substitute narrower proof unless the user accepts a narrowed claim.
Routing
- Use
../testing-best-practices/SKILL.md when durable tests are added,
changed, deleted, reviewed, or relied on as proof.
- Use
$playwright-interactive, $playwright, or $screenshot for browser,
Electron, or OS screenshot mechanics when installed.
- Use the Test Android Apps plugin for Android emulator/device QA or
performance mechanics.
- Use the security plugin for explicit security scans or finding fixes.
- Use
../user-apps-design/SKILL.md and design_judge for broad visual
quality approval.
References
- Read
references/browser-runtime-proof-workflow.md before browser-visible
runtime proof.
- Read
references/browser-proof-layering-contract.md before choosing a
browser proof channel.
- Read
references/mobile-runtime-proof-workflow.md before mobile runtime
proof.
- Read
references/mobile-emulator-proof-contract.md when emulator/device
lifecycle or parallel-device constraints matter.
