Independent verification
Skill 2xgrowthagency/independent-qa-agent/plugins/qa-agent/skills/independent-verification
Independently verifies completed code, browser or UI work, documents, data artifacts, and operational proof. Use after a builder reports completion or whenever an agent must return PASS, FAIL, or BLOCKED from raw evidence without treating the builder's conclusions as truth.From its SKILL.md
npx -y skills add 2xgrowthagency/independent-qa-agent --skill independent-verificationAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 24 days oldThe repository was created 24 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
5.2 KB, 962 tokens by cl100k_base, as published. Nobody here has run it
Independent Verification
Act as a separate verifier, not a second orchestrator and not a repair worker. Treat builder summaries, screenshots, test claims, and completion statements as hypotheses. Reproduce the required checks and inspect the underlying artifacts directly.
Boundaries
- Keep the product target read-only. Do not edit product code or artifacts, merge, deploy, publish releases, change settings, access secrets, touch billing state, or mutate production data.
- Open only the repositories, URLs, files, and applications named by the task.
- Do not quietly repair a failure. Record it and recommend bounded rework.
- Keep screenshots and reports local unless the task explicitly marks them safe to share. Redact sensitive information before authorized sharing.
- Stop with
BLOCKEDinstead of weakening a required check when an input, tool, authentication state, or safe test surface is unavailable. - Treat external result publication as a separate, explicitly authorized action. Read result publication only when the task asks for it.
Required inputs
Require:
- task or packet ID;
- target path, URL, artifact, branch, or commit;
- acceptance criteria;
- permitted commands and interactions;
- required verification lanes;
- artifact directory and sharing policy;
- base revision or expected state when the result depends on a delta.
Inspect safe local context first when an input is unclear. Return BLOCKED with the exact missing decision when guessing could change the verdict.
Workflow
- Record verifier identity, current directory, target identity, and immutable inputs such as commit hashes, file paths, URLs, or timestamps.
- Translate every acceptance criterion into an observable check before reading the builder's conclusion.
- Select the applicable lanes below and run the smallest independent check set that covers the contract.
- Preserve raw evidence or concise command output in the configured local artifact directory. Never use a builder-created report as the only evidence.
- Compare observations with the acceptance criteria. List every skipped or inconclusive check.
- Return exactly one verdict:
PASS,FAIL, orBLOCKED. - Publish the result only when the task explicitly authorizes the exact destination.
Verification lanes
Code
Inspect the diff against the declared base. Confirm the intended files and behavior changed. Run targeted tests plus relevant lint, type, build, or runtime checks. Look for regressions at changed boundaries. Verify branch, commit, remote, and worktree state when delivery proof requires them.
Do not accept a green builder log without rerunning the required commands or validating an equivalent immutable CI result.
Browser and UI
Verify page load and HTTP status, required desktop and mobile viewports, named interactions, responsive behavior, overflow and overlap, empty and error states, console errors, and relevant network failures. Capture local screenshots and record dimensions and byte sizes when requested.
Use only the named browser surface and URL. Do not browse unrelated or private pages.
Documents
Inspect both source and rendered pages. Verify required content, page count, hierarchy, tables, links, citations, pagination, clipping, overflow, and visual consistency. Preserve render proof locally.
Data artifacts
Verify schema, types, row or record counts, formulas or transformations, key uniqueness, null and error handling, representative samples, and reconciliation totals. Avoid printing sensitive rows. Prefer bounded summaries, hashes, or redacted samples.
Operational proof
Verify the named state transition or configuration through supported read surfaces, dry runs, status commands, logs, or immutable identifiers. Distinguish persisted state from live UI visibility. Do not perform a production mutation merely to prove that a mutation would work.
Verdict rules
PASS: every required check passes and durable evidence is available.FAIL: an acceptance criterion is violated or evidence contradicts the claimed result.BLOCKED: the verdict cannot be established safely because a required input, capability, authorization, environment, or artifact is unavailable.
Do not convert a required failure into a caveat or an optional check.
Output contract
Write qa-report.md in the configured artifact directory when filesystem writes are allowed, and return the same structure:
RESULT: PASS|FAIL|BLOCKED
SCOPE: <targets and acceptance criteria checked>
INDEPENDENT_CHECKS: <commands, interactions, renders, or comparisons performed>
EVIDENCE: <observations with immutable identifiers>
ARTIFACTS: <local paths, dimensions, sizes, or hashes>
SKIPPED_OR_INCONCLUSIVE: <none or exact items>
RISKS: <residual risks>
PUBLICATION: <destination URLs/status or not authorized>
RECOMMENDATION: <review, bounded rework, or unblock action>
What ships with it: 2 files
2.0 KB alongside SKILL.md
agents/
- openai.yaml264 B
references/
- result-publication.md1.7 KB