Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Ralph-Style Autonomous Development Skill · majiayu000 | DecimalAI

majiayu000/ralph-autonomous-dev

SkillScore82Maintained

Enables autonomous development loops that run until all tasks pass. Use when the user says "until done", "keep going", "finish this", "終わるまでやれ", or requests long-running autonomous iteration with fix_plan.md tracking.

agent-toolingby majiayu000

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

Coding pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/majiayu000-ralph-autonomous-dev/SKILL.md

decimalai skills pull majiayu000-ralph-autonomous-dev --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/majiayu000-ralph-autonomous-dev/SKILL.md https://app.decimal.ai/s/majiayu000-ralph-autonomous-dev/SKILL.md

Raw file: /s/majiayu000-ralph-autonomous-dev/SKILL.md ↗ · machine-readable summary at /s/majiayu000-ralph-autonomous-dev.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 25d ago

—+32% tokens0% turns82% of whole cases pass with skill (18/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-07	✗→✓	▲ Improved	19%	0%
case-08	✗→✓	▲ Improved	0%	0%
case-12	✗→✓	▲ Improved	17%	0%
case-13	✗→✓	▲ Improved	18%	0%
case-14	✗→✓	▲ Improved	60%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/10/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

majiayu000 / claude-skill-registry

skills/agent/ralph-autonomous-dev-daisuke134-anicca-products/SKILL.md

View on GitHub ↗

License MITImported 8/10/2026Version v1

Skill contents

SKILL.md

Ralph-Style Autonomous Development Skill

When to Use This Skill

Invoke this skill when:

User says "until done", "keep going until complete", "finish this"
Task requires multiple iterations (tests passing, all features complete)
Long-running autonomous development is requested
User explicitly mentions "Ralph" or "autonomous loop"

Core Concept

Ralph is a bash wrapper for Claude Code that keeps you running until the job is done. This skill teaches you to work in a Ralph-compatible way, even without the external wrapper.

The Ralph Pattern

1. Read fix_plan.md → Pick next incomplete task
2. Implement → Run tests
3. Update fix_plan.md → Mark complete or note blockers
4. Repeat until ALL tasks complete
5. Output EXIT_SIGNAL: true when truly done

File Structure

Maintain these files in .ralph/:

| File | Purpose | |------|---------| | PROMPT.md | High-level project goals and rules | | fix_plan.md | Checkbox task list (current work) | | AGENT.md | Build/test commands |

fix_plan.md Format

markdown
# Current Sprint Tasks

## Priority 1 (Critical)
- [ ] Task description here
- [x] Completed task

## Priority 2 (Important)
- [ ] Another task

## Blockers
- None currently

Working Loop Behavior

Before Each Implementation

Read fix_plan.md
Select the first unchecked - [ ] item
Announce: "Working on: task name]"

After Each Implementation

Run tests (use commands from AGENT.md)
If PASS: Mark task - [x] in fix_plan.md
If FAIL: Fix the issue, repeat
Move to next task

Completion Detection

DO NOT output EXIT_SIGNAL: true unless:

ALL checkboxes in fix_plan.md are [x]
ALL tests pass
No remaining work

Output format when truly complete:

RALPH_STATUS:
  tasks_complete: true
  tests_passing: true
  EXIT_SIGNAL: true

Circuit Breaker Rules

Stop and ask for help if:

Same error 3+ times in a row
No progress for 3+ iterations
Blocked by external dependency (API key, permission, etc.)

Example Workflow

User: "Implement feature X until all tests pass"

Claude Code:
1. Check/create .ralph/fix_plan.md
2. Break feature X into subtasks
3. Loop:
   - Pick first unchecked task
   - Implement
   - Run tests
   - Mark complete if passing
   - Repeat
4. When all done:
   RALPH_STATUS:
     tasks_complete: true
     tests_passing: true
     EXIT_SIGNAL: true

Integration with CLAUDE.md

Add to project's CLAUDE.md:

markdown
### Autonomous Development
- For "until done" tasks, follow `.claude/skills/ralph-autonomous-dev/SKILL.md`
- Always use fix_plan.md to track progress
- Output EXIT_SIGNAL: true only when truly complete

Commands Reference

bash
# Build & Test iOS
cd aniccaios && fastlane test

# Build & Test API
cd apps/api && npm test

# Run E2E Tests
maestro test maestro/ --include-tags smokeTest

Key Principles

Never exit prematurely - Only EXIT_SIGNAL when ALL work is done
Track progress explicitly - Update fix_plan.md after each task
Test after every change - Don't mark complete without passing tests
Ask for help when stuck - Circuit breaker prevents infinite loops
Incremental commits - Commit after each completed task

Common Patterns

Pattern: Feature Implementation

1. Create fix_plan.md with subtasks
2. Implement each subtask
3. Test after each
4. Commit when task passes
5. Move to next

Pattern: Bug Fix

1. Reproduce the bug (write failing test)
2. Fix the code
3. Verify test passes
4. Commit
5. EXIT_SIGNAL: true

Pattern: Refactoring

1. Ensure all tests pass (baseline)
2. Make incremental changes
3. Run tests after each change
4. Commit when stable
5. EXIT_SIGNAL: true when complete

References

Illustrative Want to see what majiayu000/ralph-autonomous-dev would put in front of your model? Preview the rendered body and check its trigger phrases — no LLM call, no signup.

Latest benchmark · publisher runVerified

9/3/2026, 04:18 UTC · 22 cases

Measured on 19 comparable cases — below the 22 needed to publish a lift, so this result is shown here but not on the registry card or leaderboard.

Lift

+32%

82% vs 50%

Agent turns

1.0 vs 1.0

Tokens

+32%

1,950.89 vs 1,478.32

fail → pass · 7pass → pass · 11fail → fail · 4

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 I need you to refactor our authentication service to use JWT tokens instead of session cookies. Keep working autonomously on this until complete. Please break down the refactoring into items on a tracking checklist, work through each item one by one while running tests after every modification, and summarize the final completion status block once all tests pass and everything is finished.	fail→fail	2,491	15,692	+530%	1	1	0%	285	1,357	+376%	0	0	—
case-02 Please fix the reported memory leak in our API server and keep going until all tasks are finished. Set up a step-by-step fix plan, implement the fixes incrementally, verify each change against our test suite, and present a structured status report confirming full task completion and passing tests when you are completely done.	fail→fail	8,509	15,058	+77%	1	1	0%	216	1,328	+515%	0	0	—
case-03 Execute the Ralph autonomous loop to implement user profile picture uploads. Work continuously until the entire feature is done, maintaining a task list where you mark off completed items, running the test command after each change, and outputting the final termination status signal once all checks pass.	fail→fail	14,798	9,827	-34%	1	1	0%	238	1,328	+458%	0	0	—
case-04 Can you explain how JavaScript closures work in simple terms with a 5-line code snippet?	pass→pass	9,384	12,007	+28%	1	1	0%	1,043	2,471	+137%	0	0	—
case-05 Review this git diff for potential race conditions: `if (!user) { await createUser(); }` and suggest improvements.	pass→pass	15,613	10,600	-32%	1	1	0%	1,935	3,038	+57%	0	0	—
case-06 Reformat this JSON object to use 2-space indentation and sorted keys: `{"b": 1, "a": 2}`.	pass→pass	1,480	2,966	+100%	1	1	0%	254	1,443	+468%	0	0	—
case-07 We want to set up an autonomous build tracking file for our new search module feature until done. What specific directory and filename should store the main task checklist for this workflow?	fail→pass	7,718	2,549	-67%	1	1	0%	1,320	1,576	+19%	0	0	—
case-08 Where should build and test commands like `npm test` or `fastlane test` be documented for the autonomous iteration agent to reference?	fail→pass	13,270	8,057	-39%	1	1	0%	2,165	2,158	-0%	0	0	—
case-09 In which configuration file under the Ralph state directory should high-level project goals and rules be stored?	pass→pass	16,805	1,656	-90%	1	1	0%	1,876	1,353	-28%	0	0	—
case-10 Before editing any files to tackle the first pending item in the queue during an autonomous run, what explicit string pattern must be announced in the output?	fail→fail	27,232	7,069	-74%	1	1	0%	2,199	1,369	-38%	0	0	—
case-11 During an ongoing autonomous run, running `npm test` throws the exact same missing native library linking error on iteration 1, iteration 2, and iteration 3. What action should be taken?	pass→pass	13,473	9,697	-28%	1	1	0%	1,474	1,869	+27%	0	0	—
case-12 You have completed 3 iteration loops on a complex refactor without making any net task progress or passing new tests. How should the autonomous worker react?	fail→pass	18,121	6,722	-63%	1	1	0%	1,854	2,173	+17%	0	0	—
case-13 While attempting to execute integration tests autonomously, the test runner fails because an `STRIPE_SECRET_KEY` environment variable is not set on the local system. What should happen next?	fail→pass	10,317	10,309	-0%	1	1	0%	1,681	1,988	+18%	0	0	—
case-14 All tasks in the plan are checked off as complete and all test suites pass. Write the complete structured status block that signals execution termination.	fail→pass	5,745	7,692	+34%	1	1	0%	854	1,367	+60%	0	0	—
case-15 All task items in `fix_plan.md` have been marked complete `- [x]`, but running `fastlane test` results in 1 failing test case. Should `EXIT_SIGNAL: true` be emitted?	pass→pass	7,950	4,398	-45%	1	1	0%	895	1,934	+116%	0	0	—
case-16 All integration tests pass, but there is still one `- [ ]` task remaining in `fix_plan.md`. Can the autonomous worker emit `EXIT_SIGNAL: true`?	pass→pass	14,660	8,938	-39%	1	1	0%	1,291	1,850	+43%	0	0	—
case-17 You need to fix a null pointer bug in the user login pipeline using the Ralph methodology. What is the very first step before modifying source code?	pass→pass	6,541	10,267	+57%	1	1	0%	915	1,995	+118%	0	0	—
case-18 We are starting a multi-file database layer refactor autonomously. What must be established before making any structural code modifications?	pass→pass	20,693	12,704	-39%	1	1	0%	2,014	2,290	+14%	0	0	—
case-19 A subtask item was `- [ ] Add input validation`. The tests just passed. How should this item be represented in `fix_plan.md` now?	pass→pass	9,625	3,011	-69%	1	1	0%	780	1,551	+99%	0	0	—
case-20 When configuring a repository's `CLAUDE.md` to instruct future AI sessions on how to perform autonomous loops in this codebase, what rule regarding progress tracking must be included?	fail→pass	13,242	8,993	-32%	1	1	0%	1,282	1,791	+40%	0	0	—
case-21 In an autonomous loop with 4 separate tasks in `fix_plan.md`, when should git commits be executed?	pass→pass	17,706	11,118	-37%	1	1	0%	1,938	2,221	+15%	0	0	—
case-22 Draft a template for `fix_plan.md` showing how critical tasks and external obstacles should be structured.	fail→pass	19,841	15,663	-21%	1	1	0%	2,318	2,630	+13%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted, and 19 counted toward the lift figure. The other 3 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +32 percentage points is the difference between those two pass rates over the 19 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 50%
Pass rate with the skill: 81.8%
Cases attempted: 22
Cases counted toward lift: 19
Cases that regressed: 0
Skill version measured: v1
Run completed: 9/3/2026, 04:18 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-07

Prompt

We want to set up an autonomous build tracking file for our new search module feature until done. What specific directory and filename should store the main task checklist for this workflow?

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run82

Live evalNeeds ≥10 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns+53% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 24 days ago

Other measured skills in the registry, with their headline benchmark lift.