Guide for using /ralph-loop command for long-running autonomous tasks. Use when starting iterative development loops, implementing Bedrock plugins, or running test-fix cycles. (project)
Guide users on effectively using /ralph-loop for long-running autonomous tasks in the mineflayer-bedrock project.
The Ralph Wiggum plugin implements an iterative AI development loop. It uses a Stop hook that intercepts Claude's exit attempts, feeding the same prompt back to continue working until completion criteria are met.
Core mechanism: You run /ralph-loop "task" once. Claude works, tries to exit, the hook blocks it, re-feeds the prompt, and repeats until done.
CLI arguments cannot contain newlines due to Bash security restrictions. These GitHub issues track this:
Workaround: Keep prompts single-line or use phased execution (see below).
--completion-promise uses exact string matching. You cannot use regex or multiple conditions like "SUCCESS OR BLOCKED". Use the <promise>TAG</promise> pattern for clarity.
Each iteration consumes full context window tokens. A 50-iteration loop on a large codebase can cost $50-100+ in API credits.
This is your primary safety mechanism. Never rely solely on completion promises.
/ralph-loop "Your task..." --max-iterations 30
Wrap completion signals in tags for clear detection:
/ralph-loop "Build feature X. Output <promise>COMPLETE</promise> when done." --completion-promise "COMPLETE" --max-iterations 25
Add escape instructions for when Claude gets stuck:
/ralph-loop "Implement feature. After 10 iterations without progress: document blockers, list attempts, suggest alternatives. Output <promise>BLOCKED</promise> if truly stuck, <promise>DONE</promise> if complete." --max-iterations 20
Break large tasks into phases instead of one massive prompt:
# Phase 1
/ralph-loop "Phase 1: Create data models. Output <promise>P1_DONE</promise>" --max-iterations 15
# Phase 2
/ralph-loop "Phase 2: Implement API endpoints. Output <promise>P2_DONE</promise>" --max-iterations 20
# Phase 3
/ralph-loop "Phase 3: Add tests, ensure all pass. Output <promise>P3_DONE</promise>" --max-iterations 25
/ralph-loop "Implement inventory.mts Bedrock plugin in lib/bedrockPlugins/ matching the Java API in lib/plugins/inventory.js. Map Bedrock packets to same bot behavior. Run 'npm run mocha_test' after changes. Output <promise>PLUGIN_COMPLETE</promise> when tests pass." --max-iterations 30
/ralph-loop "Run 'npm run test:bedrock'. Fix any failing tests. Do not skip tests. Output <promise>TESTS_PASS</promise> when all tests green." --max-iterations 20
/ralph-loop "Compare lib/bedrockPlugins/health.mts with lib/plugins/health.js. Ensure Bedrock version exposes same API (functions, events, properties). Fix discrepancies. Run tests. Output <promise>API_MATCH</promise> when compatible." --max-iterations 15
/ralph-loop "Run 'npx tsc --noEmit'. Fix all TypeScript errors. Do not use 'any' type. Output <promise>TYPES_CLEAN</promise> when no errors." --max-iterations 25
| Iterations | Estimated Cost |
|---|---|
| 10-20 | $10-25 |
| 30-50 | $30-75 |
| 50+ | $50-150+ |
Tips to reduce costs:
--max-iterations and increase if neededStart an iterative development loop.
/ralph-loop "<prompt>" --max-iterations <n> --completion-promise "<text>"
Options:
--max-iterations <n>: Stop after N iterations (required for safety)--completion-promise <text>: Exact string that signals completionCancel an active Ralph loop immediately.
/cancel-ralph
| Mistake | Prevention |
|---|---|
| No iteration limit | Always use --max-iterations |
| Vague completion criteria | Use specific, testable conditions |
| Multiline prompts | Keep single-line or use phases |
| Relying only on promise | Set iteration limit as backup |
| Too many iterations at once | Start small, increase if needed |
skill-builder: For creating new Claude Code skillscreate-analyzer: For packet analysis tasks