Skip to content
crizCraigPublic

About

Agent skills for Claude Code and Codex: implement → independent review → fix loops

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

17 Commits

Folders and files

Repository files navigation

skills

Personal skills library, consumed by two different CLIs that expect two different layouts.

Index

Skill What it does
ralph-review Iterative in-session review-and-fix loop over the current Git changes.
agent-bus-review Implement -> independent reviewer (Codex CLI) -> fix -> quick-final re-check loop, coordinated through a file-based agent bus; scripts, prompt templates and a bus template included.

Claude Code — flat <name>.md

Claude Code loads skills as flat Markdown files (frontmatter + prompt body), one per skill, e.g. ralph-review.md. These are wired up via symlinks from ~/.claude/commands/<name>.md -> ~/src/skills/<name>.md, which is what makes /ralph-review available as a slash command in Claude Code sessions.

Codex — <name>/SKILL.md with optional agents/openai.yaml

Codex discovers user-wide skills under ~/.agents/skills. Each skill is a subdirectory containing:

  • SKILL.md — same frontmatter + prompt body idea, but with Codex-oriented wording (e.g. it explicitly notes git-mutating side effects and gates invocation on explicit user request/authorization).
  • agents/openai.yaml — Codex-specific interface/policy config (interface.display_name, interface.short_description, interface.default_prompt, policy.allow_implicit_invocation, etc.).

Install the Ralph Review skill for the current user with a symlink so edits in this repository are picked up automatically:

mkdir -p ~/.agents/skills
ln -s ~/src/skills/ralph-review ~/.agents/skills/ralph-review

Verify the link and skill manifest before starting a new session:

readlink ~/.agents/skills/ralph-review
test -f ~/.agents/skills/ralph-review/SKILL.md

Do not install personal skills under ~/.codex/skills; it is not a documented user discovery location. Codex follows skill-directory symlinks and normally detects changes automatically. Restart Codex if a new or updated skill does not appear, then use /skills to confirm it is available. Codex may omit skills from the initial prompt's abbreviated catalog when many skills are installed, but explicit invocation remains available.

See the official OpenAI documentation for the current discovery locations and symlink behavior: https://learn.chatgpt.com/docs/build-skills#where-codex-loads-local-skills.

Ralph Review disables implicit invocation because it mutates Git state. Invoke it explicitly as $ralph-review.

For ralph-review, both formats currently exist side by side (ralph-review.md for Claude Code, ralph-review/ for Codex) and are maintained independently — edits to one do NOT propagate to the other. Keep both in sync by hand when changing the skill's behavior.

agent-bus-review

A single folder (agent-bus-review/) used by both CLIs; no separate flat file. Install by symlinking the folder:

mkdir -p ~/.claude/skills ~/.agents/skills
ln -s "$PWD/agent-bus-review" ~/.claude/skills/agent-bus-review   # Claude Code
ln -s "$PWD/agent-bus-review" ~/.agents/skills/agent-bus-review   # Codex

Then create the bus once from agent-bus-review/bus-template/ (see the skill's Setup section). Scripts need bash, codex, and optionally fswatch/inotifywait.

About

Agent skills for Claude Code and Codex: implement → independent review → fix loops

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages