Qa Gate

Standardized QA process with 8 checks and lobster debate logic to ensure artifacts are accurate, complete, consistent, sound, and free of sensitive data befo...

QA Gate · 质量检测门

龙虾兵团专属质量门。每个交付物在交付前必须经过此门检测。 for any artifact before human review. Every document, skill, blog post, PRD, or code output should pass this gate before the principal sees it.

This is not a code review skill. It is a read-only release gate that determines whether an artifact is ready to move forward. QA Gate inspects artifacts but does not modify them.

When to Use

• After any ralphy loop completes a PRD
• Before presenting any deliverable to the principal
• When self-reviewing documents, code, skills, or blog posts
• As the final step before publishing to ClawHub or Gumroad
• When asked to "QA gate this," "validate before publish," "final check," or run a "quality gate"

Optional Mode

• --dual: Use cross-model QA validation when the artifact is high-stakes, ambiguous, or worth the extra cost/latency for a second independent quality pass.

Process

Step 1: Read the artifact completely

Read the entire file. Do not skim. Understand the structure, voice, and intent.

Step 2: Validate against 6 dimensions

1. Factual Accuracy (Sequential Claim Verification) Extract every verifiable claim from the artifact into a mental checklist. Then verify each independently — do not batch-assess. For each claim:

• Is it verifiable from a known source or self-evident from context?
• If it references a citation (paper title, arXiv ID, finding), does the citation match?
• If it describes a technical procedure, is the procedure feasible as described?
• If it references a tool, API, or version, is the reference accurate and current?

Score: count of verified claims / total claims. If verification rate < 90%, flag for revision.

2. Tone & Voice Consistency

• Does the document maintain its intended voice throughout?
• No tonal drift between sections?
• No marketing fluff, tutorial-speak, or filler?
• Appropriate for the target audience (agent, human, or both)?

3. Completeness

• No placeholders (, TBD, 📝 计划修复, PLACEHOLDER, [FILL IN])?
• All sections referenced in TOC/structure are present?
• All promised content is delivered?
• No orphaned references or dead links?

4. Structural Integrity

• Heading hierarchy is clean (no skipped levels)?
• Code blocks are properly fenced and syntactically valid?
• Section anchors work?
• Back-links resolve to valid targets?
• Markdown renders correctly?

5. Operational Soundness (for technical documents)

• Procedures are implementable as described?
• Configuration formats match the actual system?
• Commands and scripts are executable?
• Edge cases are addressed?

6. Sensitive Data Check

• No personal information (real names, schedules, addresses)?
• No API keys, tokens, or secrets?
• No internal-only references that shouldn't be public?
• Examples use fictional/generic data?

Step 3: Produce gate verdict

Output must include a clear gate result:

PASS — ready for human review

or

PASS WITH FIXES
- MINOR [location]: issue description

or

FAIL
- CRITICAL [location]: issue description
- MAJOR [location]: issue description
- MINOR [location]: issue description

Step 4: If FAIL, fix and re-validate

Fix all CRITICAL and MAJOR issues. Re-run the gate. Only present to principal after PASS or PASS WITH FIXES.

Integration with PRD Workflows

Add to any PRD as a verification step:

### D) QA Gate
- [ ] Run QA Gate on all major artifacts produced in this PRD
- [ ] All artifacts must PASS before marking PRD complete
- [ ] Fix any CRITICAL or MAJOR issues identified

Output Format

Write validation report to: qa-gate/YYYY-MM-DD-<artifact-slug>.md (relative to your workspace or evidence directory)

Use this structure:

# QA Gate Report: <artifact name>

## Gate Result
PASS | PASS WITH FIXES | FAIL

## Artifact Type
Document | Skill | PRD | Blog Post | Code Artifact | Other

## Findings
- SEVERITY [location]: issue description

## Summary
Brief explanation of why the artifact passed, passed with fixes, or failed.

Quality Standards

• CRITICAL: Blocks release. Factual errors, security issues, broken functionality.
• MAJOR: Should fix before release. Missing sections, tone drift, incomplete content.
• MINOR: Nice to fix. Typos, formatting inconsistencies, style preferences.

A PASS with only MINOR issues is acceptable. CRITICAL or MAJOR = must fix first.