Agent Skills: Goal Loop

Goal Loop

合意した Scope と success criteria を満たすまで、worker / verifier / evaluator を使ってゴールをやり遂げる手順書。 このスキルを起動した現在の agent が orchestrator になり、criteria、ledger、evidence reconciliation、次の一手の決定を担う。

When to Use

• ユーザーが goal-loop / ゴールまで回す / 必ずやり遂げる / loop until done / until it actually works と依頼した。
• 成果物を end-to-end に完成させる必要があり、単発修正ではなく検証・再試行・評価のループが必要。
• worker 生成と evaluator 判定を分ける価値がある、中規模以上または不可逆・多段の作業。

Do Not Use For

• 1 ファイル・自明・低リスクの修正。通常 agent が直接実行する。
• 普通の相談、計画だけ、コードレビューだけ、事実確認だけ。
• skill / prompt / instruction / agent 自体の設計レビュー。まず skill-creator-plus や agentic-workflow-guide を使う。

Execution Contract

• Freeze scope, AC, verifier, out-of-scope, must NOT, autonomy, and persistence before execution.
• Keep a compact ledger: AC status, evidence, attempts, gaps, and next action.
• Delegate broad research, mutation, and primary verification when workers are available; workers return evidence only.
• Judge by external signals first. Never PASS from self-evaluation, supporting checks alone, skipped tests, TODOs, or weakened criteria.
• On failure, update the ledger from the failed signal, change approach, and avoid repeating attempts.
• Use Small-Bet-First for broad, irreversible, dependency, schema, deployment, or user-facing changes.
• Stop or replan on stall, oscillation, diminishing returns, budget exhaustion, or genuine external blockers.

Phase 1: Scope + Criteria Freeze

Use criteria-agreement.md.

• Freeze goal, scope terminus, AC with verify:, primary verifier, out-of-scope, must NOT, autonomy, and persistence.
• If the finish line is ambiguous, ask one structured question before execution.
• Do not use build/control-plane success as the primary verifier when the real outcome is deploy / runtime / UI / app behavior.

Phase 2: Plan Minimal Subgoals

Use steering-and-final-gates.md and ledger-templates.md.

• Pick the lightest effort level that still protects the goal; downgrade trivial work.
• Split into verifiable subgoals, record status, parallelize only independent read-only work.

Phase 3: Delegate Workers

Use loop-control.md.

• Start broad or risky changes with a pilot; never expand a failed pilot.
• Give workers only the subgoal, AC, boundaries, must NOT, relevant attempt lessons, and expected evidence.
• If no subagent tool exists, record degraded mode; only downgraded small work may run in main context.

Phase 4: External Verification

• Run the frozen verify: steps and capture exit codes, logs, screenshots, readbacks, or capability gaps.
• If the primary verifier is unavailable, record exact manual verification steps and required evidence; do not complete from weaker checks.

Phase 5: Evaluate

Use evaluation-rubric.md.

• Reconcile evidence against each AC. Use a separate evaluator for non-small or non-read-only work.
• PASS only when every must AC passes and confidence is above threshold.

Phase 6: Loop Control

Use loop-control.md.

• If unmet, update the ledger from the external signal and replan before adding workers.
• Escalate only after the Blocker Test proves the blocker is external or structural.

Phase 7: Report or Handoff

• Report AC-by-AC status, primary evidence, supporting checks, capability gaps, criteria changes, and residual risks.
• For high-risk changes, include final quality gate and independent review evidence.
• Cleanup scratch state unless it is needed for resumability.

References

• criteria-agreement.md: Scope / criteria freeze テンプレ。
• ledger-templates.md: Task / Progress ledger と attempt table。
• loop-control.md: stop、replan、Small-Bet、reflection、HITL。
• evaluation-rubric.md: rubric と judge bias 対策。
• steering-and-final-gates.md: Effort Scaling、dynamic steering、final gate、worker boundary。
• design-rationale.md: 設計根拠と出典。