CtrlK
BlogDocsLog inGet started
Tessl Logo

uinaf/verify

Run the builder-owned pre-review verification pass for a completed change using repo guardrails and real-surface evidence. Use a separate evaluator for complex, subjective, or high-risk work; when the repo is not verifiable, report blocked with the exact missing infrastructure and required setup.

94

1.04x
Quality

95%

Does it follow best practices?

Impact

94%

1.04x

Average score across 3 eval scenarios

SecuritybySnyk

Passed

No known issues

Overview
Quality
Evals
Security
Files

SKILL.md

name:
verify
description:
Self-check an agent's own completed change before independent review — the builder-owned pre-review sanity pass. Use when running repo guardrails, exercising changed behavior end to end, collecting evidence, or catching obvious self-correctable issues. Produces a `ready for review` / `needs more work` / `blocked` verdict — never a ship decision. If the repo cannot be booted or exercised reliably, report `blocked` with the exact missing infrastructure and required setup. Do not use to audit someone else's diff, branch, or PR.

Verify

Run the builder-owned proof that a completed change boots, passes guardrails, and survives the real surface before independent review.

Principles

  • Keep proof layers separate: local guardrails, focused regression checks, real-surface runtime proof, CI, and live/deploy evidence are related but not interchangeable

Boundaries

  • This pass reports readiness for independent review; it never decides whether to ship.
  • No stable boot, smoke, or interaction path → report blocked, name every missing prerequisite, and specify the smallest durable install, boot, test, smoke, or interaction surface required.
  • Auditing existing code, a diff, branch, or PR you did not author is independent review work.
  • Stale AGENTS.md, README, specs, or repo docs outside the changed contract are out of scope; report the exact drift unless it blocks verification.

Before You Start

  1. Define the exact change being verified and the expected user-visible behavior
  2. Decide whether the change needs a separate evaluator: use one for complex, subjective, or high-risk work; otherwise keep this as a direct builder sanity pass
  3. Load the target repo's guidance files such as AGENTS.md, CLAUDE.md, or repo rules, when present
  4. Confirm you can boot and interact with the real surface
  5. Define the proof boundary: what local, CI, live, or provider-specific claim you are actually verifying
  6. Pick the smallest check set that can disprove the change honestly

Workflow

1. Run deterministic guardrails first

  • Prefer the repo's built-in entrypoint: make verify, just verify, pnpm test, cargo test, or the nearest targeted equivalent
  • During iteration, run targeted checks to avoid context flooding; before handoff, run the canonical local gate when it exists and is feasible
  • When choosing tests, prefer the strongest cheap proof available: smoke, integration, contract, or e2e checks beat mock-heavy unit suites that mainly replay implementation details
  • Swallow boring success output and surface only failures, anomalies, and exact commands

2. Exercise the real surface

  • UI → run the browser automation, navigate the changed flow, and capture screenshots
  • API → hit the local endpoint with a real request such as curl http://127.0.0.1:3000/health
  • CLI → run the shipped command such as node dist/cli.js --help or the repo's packaged entrypoint
  • state/config → verify round trips, restart behavior, and config boot paths
  • deploy/live wiring → prove the actual configured surface when required; do not imply production readiness from a local build alone

Follow references/evidence-rules.md when collecting proof.

3. Self-correct obvious issues when authorized

When the surrounding implementation or fix task authorizes edits, fix anything cheap and obvious that you spot while exercising the change. If the user asked only for verification or a report, do not edit; record the issue and return needs more work.

  • A typo in a log line, a stale comment, an unused import, a duplicated helper inside the diff
  • An any, unsafe as, or non-null assertion you can replace with a real type in seconds
  • A failure path that swallows errors silently when a one-line throw makes the diagnostic useful

Keep this as a self-check pass. Leave substantive code-shape concerns (architecture mismatches, broader duplication, error-classification redesigns) for independent review. Use references/simplification.md as a short self-check.

4. Probe adjacent risk

  • Check the main happy path
  • Check at least one failure path or edge case
  • Check that at least one exercised failure path returns or logs a useful, actionable error instead of a vague or swallowed failure
  • Re-test any config, persistence, or restart-sensitive behavior touched by the change

5. Synthesize the verdict

Produce one clear outcome:

  • ready for review — guardrails green, real surface confirmed, no obvious self-correctable issues left
  • needs more work — the change is not ready to be reviewed; specific issues to address are listed
  • blocked — verification cannot proceed, usually because infrastructure is too weak

If a requested proof surface was unavailable, name that boundary explicitly instead of substituting a weaker check.

Output

After verification, report in this compact bullet shape:

  • - verdict: exactly one of ready for review, needs more work, or blocked
  • - evidence: concise explanations of what checks proved, not full commands
  • - fixed during verify: only if self-corrections happened
  • - unverified or gaps: readiness gaps, doc drift, or none
  • - next: independent review, build missing verification infrastructure, update affected documentation, more implementation, or none

Keep the final answer short:

  • Put detailed failures, screenshots, traces, and file references in native findings or the work log, not in the footer
  • Summarize command output that already appeared in the terminal
  • Keep the footer to 5 labeled lines or fewer
  • Omit fixed during verify when nothing was corrected
  • Summarize passing checks by intent and result, for example typecheck passed for tv-vite or API smoke check returned 200; include full commands only when they failed, are needed for reproduction, or the user asks for them
  • For failures or blocked checks, include the relevant error/status line or response snippet in the report

Example:

- verdict: ready for review
- evidence: retry tests covered success and failure paths; API retry smoke returned 200
- unverified or gaps: none
- next: independent review

References

SKILL.md

tile.json