SKILL.md
Council Verification Skill
Use LLM Council's multi-model deliberation to verify work with structured, machine-actionable verdicts.
When to Use
- Verify code changes before committing
- Validate implementation against requirements
- Check documents for accuracy and completeness
- Get multi-model consensus on quality
Workflow
- Snapshot: pin the git commit SHA (reproducibility)
- Verify: call
mcp:llm-council/verify(isolated context) - Act on the verdict (structured JSON below)
- Audit: transcript via
mcp:llm-council/audit
Parameters
| Parameter | Type | Default | Description |
|---|---|---|---|
snapshot_id |
string | required | Git commit SHA for reproducibility |
rubric_focus |
string | null | Focus area: "Security", "Performance", "Accessibility" |
confidence_threshold |
float | 0.7 | Minimum confidence for PASS verdict |
tier |
string | "balanced" | Confidence tier: "quick", "balanced", "high", "reasoning" |
Tier Selection Guide
| Tier | Use When | Timeout | Max Input |
|---|---|---|---|
quick |
Fast sanity checks, small diffs | ~30s | 15K chars |
balanced |
Default. Routine verification | ~90s | 30K chars |
high |
Security-critical reviews only | ~180s | 50K chars |
reasoning |
Complex architectural decisions | ~600s | 50K chars |
Important: high/reasoning use slow frontier models. For large diffs (>20K chars), prefer balanced to avoid timeouts.
Evidence (ADR-042)
Pass pre-computed analysis from upstream tools (linters, slop detectors) via evidence: List[EvidenceItem]. Council emits evidencesummary and evidencewarnings in the response. See [references/evidence.md](references/evidence.md) for fields, limits, example.
Output Schema
{
"verdict": "pass|fail|unclear",
"confidence": 0.85,
"confidence_calibrated": 0.62,
"unclear_reason": null,
"rubric_scores": { "accuracy": 8.5, "completeness": 7.0, "...": "..." },
"blocking_issues": [...],
"rationale": "Chairman synthesis...",
"transcript_location": ".council/logs/...",
"partial": false,
"timeout_fired": false,
"completed_stages": ["stage1", "stage2", "stage3"],
"timing": { "total_elapsed_ms": 142000, "budget_utilization": 0.53, "...": "..." },
"input_metrics": { "content_chars": 32000, "tier": "high", "...": "..." },
"evidence_summary": [...],
"evidence_warnings": [...]
}If timeoutfired: true, the tier deadline was exceeded. Check completedstages for progress. timing.budget_utilization shows time used vs deadline (1.0 on timeout). See ADR-040 for full timeout semantics.
On UNCLEAR, route on unclearreason: infrafailure → check gateway + RETRY; low_confidence → accept-and-audit if no blocking issues; timeout → re-tier. Details: references/unclear-routing.md.
Rules
- One call at a time. Never fire multiple verify calls concurrently or in rapid succession. Wait for each to complete before deciding next steps.
- One call per commit. Never retry the same snapshot_id. If it fails, fix the code first.
- Act on verdicts, don't retry them:
- PASS (exitcode 0): Proceed. - FAIL (exitcode 1): Read blockingissues. Fix the code, commit, then re-verify the new snapshot. - UNCLEAR (exitcode 2): Accept and move on. Do not retry.
- Do not reduce scope and retry. Sending the same code with fewer files is still a retry.
Related Skills
council-review (code review), council-gate (CI/CD gate)