vstakhovsky Shokunin Verify Skill
Verifies output quality before returning to user.
Purpose
Final verification step to ensure output is safe, grounded, and useful.
When to Use
This skill runs automatically at the end of every review. User doesn't call it directly.
Workflow
- Check grounding — Every finding grounded in artifact?
- Check actionability — Every finding has concrete fix?
- Check specificity — No generic advice?
- Check duplicates — No duplicate findings?
- Check tone — Tone calm and direct?
- Check safety — No toxic language or invented evidence?
- Check consistency — Score matches findings?
Inputs
- Review findings
- Calculated score
- Verdict
Output Contract
If verification passes:
- Output is returned to user
If verification fails:
- Findings revised
- Score recalculated if needed
- Verification re-run
- Only when all checks pass → Output returned
Verification Checklist
- ✅ All findings reference artifact content
- ✅ All findings have location (line/section)
- ✅ All findings have concrete fix
- ✅ All findings have example (where helpful)
- ✅ No "improve clarity" without specifics
- ✅ No "add detail" without guidance
- ✅ No duplicate findings
- ✅ Tone is calm, not dramatic
- ✅ No shaming language
- ✅ No accusations of AI use
- ✅ No invented evidence
- ✅ Score aligns with findings
- ✅ Verdict matches score
Failure Modes
If check fails:
- Finding removed or revised
- Output re-verified
- Process repeats until all pass
Example
Bad Finding (fails verification):
"Requirements need improvement. Add more detail."
Good Finding (passes verification):
"Requirements section is not testable.
Location: Requirements section
Fix: Rewrite each requirement as Given/When/Then
Example: Given user on checkout, When enters payment, Then order confirmed"