メニュー
Autonomous PRD implementation loop — turns GitHub issues into shipped code using TDD, code review gates, and git worktree isolation. The execution engine for the grill-me → write-a-prd → prd-to-issues → ralph pipeline.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Deploy two-host credential-broker patterns where an AI agent runtime reaches public APIs only through a private credential broker/proxy, with HITL handling for secrets.
Create or edit standalone, high-fidelity cinematic HTML presentations from research briefs, outlines, video notes, scripts, or rough ideas. Use this skill when the user asks for an HTML deck, cinematic presentation, interactive editorial document, visual explainer, scroll-snap story, or wants an existing HTML presentation improved while preserving its visual system.
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Use when you have a rough product idea and want a complete PRD without sitting through an interactive grilling. Claude walks the full decision tree (edge cases, modules, schema, testing, security), self-answers with software-engineering best practices, streams the Q&A live so you can override, and writes the PRD locally with an option to push as a GitHub issue.
Break a plan, spec, or PRD into independently-grabbable GitHub issues using tracer-bullet vertical slices. Use when user wants to convert a plan into issues, create implementation tickets, or break down work into issues.
Turn the current conversation context into a PRD and submit it as a GitHub issue. Use when user wants to create a PRD from the current context.
| name | ralph |
| description | Autonomous PRD implementation loop — turns GitHub issues into shipped code using TDD, code review gates, and git worktree isolation. The execution engine for the grill-me → write-a-prd → prd-to-issues → ralph pipeline. |
RALPH (Repeated Autonomous Loop for PRD Handling) implements a PRD stored as GitHub issues — autonomously, with TDD, and a code review gate after every iteration.
This skill is the execution engine for a 4-step coding workflow:
| Step | Command | What It Does |
|---|---|---|
| 1 | /grill-me | Stress-test your idea with relentless questions before building |
| 2 | /to-prd | Turn the idea into an engineering spec (PRD) as a GitHub issue |
| 3 | /to-issues | Break the PRD into vertical-slice sub-issues with dependencies |
| 4 | /ralph | Implement each sub-issue autonomously with TDD + code review |
/ralph <issue-number> — Human-in-the-loop (start here).agents/skills/ralph/ralph-once.sh 42
Implements one sub-issue from PRD #42, then stops for your review. Run again for the next one. Recommended for your first few iterations.
/ralph <issue-number> afk [max-iterations] — Autonomous.agents/skills/ralph/afk-ralph.sh 42 20
Runs up to 20 iterations in a git worktree. Each iteration: implement one sub-issue, run tests, commit, run the local review gate with the Codex model and reasoning effort configured in .agents/skills/ralph/config.env, close the issue. When all sub-issues are done, pushes the branch, opens a PR, and drives CodeRabbit review.
/ralph <issue-number> afk-and-review [max-iterations] [max-review-rounds] — Autonomous + CodeRabbit loop.agents/skills/ralph/ralph-afk-and-review.sh 42 20 10
Compatibility entrypoint for the Codex-native worktree runner. Runs review with the model settings from .agents/skills/ralph/config.env.
/ralph <issue-number> afk-and-review-worktree [max-iterations] [max-review-rounds] — Autonomous + CodeRabbit loop (worktree, no Docker).agents/skills/ralph/ralph-afk-and-review-worktree.sh 42 20 10
Same behavior as afk-and-review but uses a git worktree at .agents/worktrees/ralph-prd-<N> for isolation instead of a Docker container. Codex runs on the host with cwd set to the worktree, using model settings from .agents/skills/ralph/config.env. Worktree is left in place after exit for inspection — final summary prints both the merge command and the worktree-cleanup commands.
When the user invokes this skill:
once | afk | afk-and-review | afk-and-review-worktree, default once), third is max implementation iterations (default 20), fourth is max review rounds for the two afk-and-review* modes (default 10).agents/skills/ralph/ exists at the project root.once mode (default):
.agents/skills/ralph/ralph-once.sh <issue-number>
/pr-review-loop).afk mode:
.agents/skills/ralph/afk-ralph.sh <issue-number> <max-iterations>
afk-and-review mode:
.agents/skills/ralph/ralph-afk-and-review.sh <issue-number> <max-iterations> <max-review-rounds>
afk-and-review-worktree mode (worktree-isolated, no Docker):
.agents/skills/ralph/ralph-afk-and-review-worktree.sh <issue-number> <max-iterations> <max-review-rounds>
afk-and-review* mode)gh CLIParent PRD #<number> in their bodySub-issues created by /to-issues follow this format:
## Parent PRD
#42
## What to build
Description of the work...
## Acceptance criteria
- [ ] Criterion 1
- [ ] Criterion 2
## Blocked by
- #43
/ralph (once mode). Validate the first iteration before going AFK./grill-me first. The more you think before coding, the better RALPH performs.