Gepetto Orchestrator
Gepetto = 17-step planning pipeline: Research → Interview → Spec Synthesis → Plan → External LLM Review → Section Files → Execution Files (ralph-loop prompt + PRD).
Skill location: ~/.agents/skills/gepetto/SKILL.md
Is Gepetto Worth It?
Ask before starting — gepetto is expensive in time and tokens.
Use Gepetto when:
- Non-trivial feature (multi-day work, multiple files, architectural impact)
- Requirements are fuzzy and need clarification
- You want external LLM review of the plan (Gemini + Codex)
- You'll hand off to ralph-loop or Ralphy for autonomous execution
Skip Gepetto when:
- Task is well-understood and < 1 day
- You already have a clear spec → go straight to
/gsd-plan-phaseor/ralph-loop - You just need a quick implementation →
/gsd-fastor/gsd-quick
Phase 1: Prepare the Spec File
Gepetto requires a .md spec file. If the user doesn't have one, help draft it.
A good spec file:
# [Feature Name]
## What
[One paragraph: what we're building]
## Why
[Business/technical motivation]
## Scope
- In: [what's included]
- Out: [what's explicitly excluded]
## Known constraints
[Tech stack, performance requirements, existing patterns to follow]
## Open questions
[Things you're unsure about — Gepetto's interview will resolve these]
Present the draft and ask: "Does this capture it? Save it as planning/[feature-name]-spec.md then run /gepetto @planning/[feature-name]-spec.md."
Phase 2: During Gepetto
Gepetto runs mostly autonomously — it spawns its own subagents. Your role during the run:
- Answer the interview questions (Step 6) — these are the most important input. Be specific.
- Review the integrated plan (Step 12) — gepetto pauses here. Edit
claude-plan.mdif needed. - Let external review run — Gemini + Codex reviews happen automatically.
Keep this chat context clear while gepetto runs. Resume here after Step 17.
Phase 3: After Gepetto — Choose Execution Path
When gepetto finishes, three outputs matter:
| File | Used for |
|---|---|
sections/ |
Manual implementation, section by section |
claude-ralph-loop-prompt.md |
Autonomous via /ralph-loop |
claude-ralphy-prd.md |
Autonomous via Ralphy CLI |
Advise on execution path
| Situation | Recommendation |
|---|---|
| Complex, want oversight | Manual: read sections/index.md, implement in dependency order |
| Want autonomy, have good test coverage | /ralph-loop @planning/claude-ralph-loop-prompt.md --completion-promise "ALL-SECTIONS-COMPLETE" --max-iterations 100 |
| Using Ralphy CLI | ralphy --prd planning/claude-ralphy-prd.md |
| Want GSD tracking alongside | Import plan into GSD with /gsd-import |
Combining with ralph-orchestrator
If using ralph-loop for execution, hand off to ralph-orchestrator skill for:
- Prompt quality review before launching
- Monitoring for stuck loops
- Windows hook fix if needed
Status Header
Show at every response:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Gepetto | Stage: [spec prep / running / post-planning]
| Spec: [filename or "not created yet"]
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Resuming a Gepetto Session
Gepetto auto-detects existing files and resumes from the right step. If the user wants to resume:
/gepetto @planning/[feature-name]-spec.md
It will scan for claude-research.md, claude-interview.md, etc. and pick up where it left off.