菜单
Continuous autonomous iteration patterns (Ralph Loops) for Claude Code. Enables self-directed build-test-fix cycles, progressive refinement, and autonomous quality improvement. Use when the user wants Claude to iterate autonomously on a task until completion.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
AI tool delegation patterns — Gemini-CLI, Codex, and other external AI tools. Apply when delegating tasks to external AI tools to save tokens.
Autonomous execution mode for GIR — unattended operation, cap fail-safe, checkpoint/resume protocol. On-demand skill loaded by /gir:run.
Universal Claude Code best practices for any coding task. Covers tool usage patterns, anti-patterns, code review standards, commit conventions, and workflow quality standards. Apply to ALL software engineering tasks.
GIR orchestration core — routing rules, delegation tiers, cost discipline, Graphify integration, escalation trigger, and lazy-load instructions. Always loaded.
Native parallel agent execution using Claude Code's built-in Agent tool with worktree isolation. On-demand skill — load when parallel work is needed.
Spec-driven quality gates (SpecKit) for Claude Code. Enforces specification verification before implementation, test coverage requirements, and acceptance criteria checks. Use when building features that require formal specifications or quality assurance workflows.
| name | ralph-loops |
| description | Continuous autonomous iteration patterns (Ralph Loops) for Claude Code. Enables self-directed build-test-fix cycles, progressive refinement, and autonomous quality improvement. Use when the user wants Claude to iterate autonomously on a task until completion. |
Continuous AI iteration until completion criteria met
Self-referential feedback loops where Claude:
Named after Ralph Wiggum — persistent iteration despite setbacks.
claude code plugins install anthropics/claude-code/plugins/ralph-wiggum
/ralph-loop "Implement feature X.
Completion criteria:
- Tests pass (npm test)
- Build succeeds (npm run build)
Output <promise>COMPLETE</promise> when done."
--completion-promise "COMPLETE"
--max-iterations 20
/cancel-ralph # If needed
I need you to work iteratively on this until completion.
[TASK]
Completion Criteria (verify each):
- [Criterion 1] (verify: [command])
- [Criterion 2] (verify: [command])
Process:
1. Implement changes
2. Run verification
3. If not passing, debug and iterate
4. If passing: check `.gir/DOD.md` if it exists — all items must pass — then output <promise>COMPLETE</promise>
Use TodoWrite to track iterations.
Keep iterating without permission until:
- All criteria met, OR
- 20 iterations (document blockers)
Start now.
See templates/RALPH-loop-prompts.md.template
| Template | Use Case |
|---|---|
| TDD Implementation | Test-driven development |
| Build Fix | Compilation/type errors |
| API Implementation | REST API with CRUD + tests |
| Refactor Until Quality | Quality metrics |
| Migration | Library/framework upgrades |
| Feature Until Deployed | End-to-end delivery |
Good for: TDD, build fixes, API implementations, long-running features (4-12h), migrations, quality improvements
Not for: Frequent design decisions, ambiguous requirements, production debugging, unpredictable iterations (>40)
Bad: Implement authentication
Good:
Implement JWT authentication
Criteria:
- Login endpoint works (POST /api/auth/login)
- All tests pass (npm test -- auth)
- Build succeeds (npm run build)
npm test # Tests
npm run build # Build
npm run type-check # TypeScript
npm run lint # Quality
Before emitting any promise, check .gir/DOD.md if it exists. All DOD items must pass — do not emit COMPLETE if they do not.
<promise>COMPLETE</promise>
<promise>TESTS_PASSING</promise>
<promise>BUILD_CLEAN</promise>
<promise>API_READY</promise>
<promise>DEPLOYED</promise>
If stuck after 15 iterations:
1. Document blockers
2. List what was tried
3. Suggest alternatives
4. Output: <promise>BLOCKED</promise>
Phase 1: Explore — Understand codebase
Phase 2: Plan — Define criteria, set verification
Phase 3: Code — RALPH LOOP HERE (iterate until complete)
Phase 4: Commit — Review (already verified), commit, deploy
| Metric | Result |
|---|---|
| Success Rate | 6 repos overnight (YC hackathon) |
| Cost Efficiency | $50k contract for $297 API |
| Long-running | Programming language over 3 months |
| Typical | 5-20 iterations for medium features |
Not completing: Break task smaller, make criteria specific, check commands, increase iterations
Exits early: Use unique promise tag, explicit criteria, add verification commands
Stuck on same error: Add fallback instructions, lower max iterations, switch to manual
Clear completion criteria? → No → Define first
Can verify automatically? → No → Don't use Ralph
Well-scoped (<40 iterations)? → No → Break into smaller
Benefits from autonomous iteration? → Yes → USE RALPH
/gir:routing-stub — Ralph as Tier 2 Recommended (see Delegation Tiers)/gir:load-workflows — Workflow detailsUpdated: 2026-01-12 | Status: Production-Ready