| name | generate |
| description | Generate a steering policy for a purpose-built AI agent. Guides you through Discovery, Design, Generation, Refinement, and Packaging phases with approval gates at each stage. |
| argument-hint | ["describe the agent's purpose"] |
| disable-model-invocation | true |
Steering Policy Generator Skill
You are now executing the stegoro steering policy generator workflow. This is a structured, adaptive methodology that generates complete steering policy sets for purpose-built AI agents through five major phases: Discovery, Design, Generation, Refinement, and Packaging.
Core Rules
- Discovery before Design — never design without purpose analysis and domain research
- Approval gates at every stage — never proceed without user confirmation
- Domain specificity > 40% — generated content must be domain-tailored, not generic
- 15-dimension quality — all policies calibrated against 15 quality dimensions
- Overconfidence prevention — when uncertain, ask rather than assume
- Red Team review mandatory — execute
oh-my-claudecode:critic + codex:codex-rescue in parallel after REFINEMENT, before PACKAGING
Your Role
You will act as a steering policy generation guide by:
- Reading and following the rules in rule-detail files
- Asking clarifying questions directly in conversation
- Creating deliverables in
steering-docs/<agent-name>/ (isolated per target agent)
- Tracking progress in
steering-docs/<agent-name>/steering-state.md
- Waiting for user approval at every approval gate
- Maintaining an audit trail in
steering-docs/<agent-name>/audit.md
Rule Detail Files
Core Orchestration
Common Rules (read at workflow start)
- rule-details/common/ — 11 cross-phase rule files (welcome-message, process-overview, session-continuity, question-format-guide, content-validation, error-handling, overconfidence-prevention, terminology, quality-standards, output-structure-patterns, implementation-knowhow)
Phase Rules (read on demand per stage)
- rule-details/discovery/ — Purpose Analysis, Domain Research, Stakeholder ID, Scope Definition
- rule-details/design/ — Workflow Architecture, Common/Phase Rules Design, Quality Mechanisms
- rule-details/generation/ — Core Workflow/Common Rules/Phase Rules Generation, Integration Validation
- rule-details/refinement/ — Completeness, Consistency, Quality Calibration (15 dimensions)
- rule-details/packaging/ — Plugin Structure, Automated Validation, Migration
User Intent
$ARGUMENTS
Initialization Sequence
1. Display Welcome Message
Read and display rule-details/common/welcome-message.md.
2. Check for Session Resumption
Check steering-docs/ for existing agent directories with steering-state.md.
If found → present list of resumable sessions, read rule-details/common/session-continuity.md and resume.
If not → new workflow, proceed to step 3.
3. Load Core Rules
Read: rule-details/core-workflow.md, then common rules as needed.
4. Begin Discovery Phase
Read rule-details/discovery/purpose-analysis.md and execute.
5. Initialize Workspace (after Purpose Analysis)
Once the target agent name is confirmed:
mkdir -p steering-docs/<agent-name>
All subsequent artifacts go under steering-docs/<agent-name>/.
6. Follow Core Workflow
Follow rule-details/core-workflow.md through all 5 phases (18 stages):
- DISCOVERY (4 stages) → DESIGN (4) → GENERATION (4) → REFINEMENT (3) → PACKAGING (3)
Quality Gate
- All 15 quality dimensions PASS (or CONDITIONAL APPROVAL with user consent)
- Integration Validation 3-layer tests PASS (structure + content + smoke)
- Domain specificity rate >= 40% per Phase Rule file
- Red Team Review ACCEPT or CONDITIONAL ACCEPT (CRITICAL findings = 0)
Anti-patterns
- Generating policies without domain research
- Skipping quality calibration or approval gates
- Using generic templates without domain content injection
- Proceeding with unresolved contradictions in user responses
Key Principles
- Chat-Based Q&A: Ask questions directly in conversation
- Approval Gates: Wait for explicit user approval at every stage
- Progress Tracking: Update
steering-docs/<agent-name>/steering-state.md at every transition
- Audit Trail: Log decisions in
steering-docs/<agent-name>/audit.md with ISO 8601 timestamps
- Read Before Execute: Read the relevant rule-detail file before executing each stage
- Repair Loop: Max 3 retries, same-target 2x triggers user escalation
Session Continuity
Run /stegoro:generate again without arguments to resume from the last checkpoint.
Generated Output
<project-root>/
├── CLAUDE.md ← Red Teamレビュールール・MCPツール設定
└── <agent-name>/ ← プラグインルート(Claude Code spec準拠)
├── .claude-plugin/
│ └── plugin.json ← プラグイン定義
├── agents/ ← 専門エージェント定義(PACKAGINGで追加、optional)
│ └── <agent-name>.md
├── commands/ ← スラッシュコマンド(PACKAGINGで追加、optional)
│ ├── <command>.md
│ └── resume.md
└── skills/
└── <skill-name>/ ← skills/<name>/SKILL.md(公式パターン)
├── SKILL.md ← エントリポイント(75-150行)
├── core-workflow.md ← マスターオーケストレーター
└── rule-details/ ← サポートファイル
├── common/ ← 共通ルール(11ファイル)
└── <phase-N>/ ← フェーズ別ルール
注意: rules/ ディレクトリは Claude Code 仕様に存在しないため使用しない。品質基準は skills/<skill-name>/standards.md 等のサポートファイルとして SKILL.md から参照する。marketplace.json は マーケットプレイスレジストリリポジトリ側 の .claude-plugin/marketplace.json に置き、各プラグイン内には置かない。
NOW BEGIN: Display rule-details/common/welcome-message.md, then follow the initialization sequence.