메뉴
Resolve ambiguous triage outcomes in the middle confidence band.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Orchestration skill: chains the full ADLC Build Loop. PRD → Brief → Council → Scaffold → Codegen → LDD → TDD → Council → PR. Use when implementing a new feature end-to-end.
Generate a bounded, schema-valid goal prompt from one validated Build Brief task for zero-context external execution.
Applicability-aware Definition of Done checklist. Core checks apply to every task; overlay checks activate only when the applicability manifest says the surface exists. Binary verification — pass or fail. Triggers at Phase 4 completion and Phase 5 entry.
Score whether each task is specific enough for autonomous, one-shot, production-ready execution.
Reviews code changes for understanding, blast radius, state, secrets, assumptions, and explainability. Produces a blocking comprehension artifact when implications are unclear.
Self-improving skill system. Captures diffs between agent output and human edits, distills patterns into rules, writes them back to skill files. Makes Build and Fix loops better over time.
| name | brief-clarity-judge |
| description | Resolve ambiguous triage outcomes in the middle confidence band. |
| contract_version | 1.0.0 |
| side_effect_profile | read_only |
| activation | {"mode":"judgement","consumes_manifest":true,"model_class":"fast_judge","cost_guard":{"max_tokens_per_call":700,"expected_calls_per_run":1}} |
Use this judge when 0.6 <= task_classification_confidence < 0.8, when epistemic_ledger contains architecture-affecting UNKNOWN entries, or when bin/adlc clarity-gate emits pending questions.
The confidence band is advisory. This judge decides whether the brief is clear enough to continue or should escalate for human clarification.
{
"request": "string",
"signal_features": {
"language_hints": ["py"],
"intent_keywords": ["fix", "bug"],
"linked_refs": ["PR-123"],
"reproducer_present": true
},
"task_classification": "bugfix",
"task_classification_confidence": 0.72,
"classification_evidence": ["..."],
"epistemic_ledger": {},
"pending_questions": [],
"contamination": {},
"missing": ["..."]
}
{
"verdict": "proceed | escalate",
"rationale": "string",
"pending_questions": [
{
"ledger_entry_id": "L-1",
"question": "string",
"why_it_matters": "string",
"what_changes": "string",
"conservative_default": "string"
}
]
}
proceed when the task has a concrete objective, bounded scope, and a plausible verifier path.escalate when the task is internally contradictory, underspecified, or missing the information needed to classify safely.escalate when an ask-user ledger entry lacks a recorded human answer or when pending questions would change architecture, compatibility, or task boundaries.escalate when a proof, demonstration, slice, or end-to-end criterion lacks a substance floor; unstated proof magnitude is an architecture-affecting unknown.ask-user unknown into an assumption; only a human answer or signed accepted risk resolves it.