From claude-codex
Use this skill when the user explicitly routes work to Codex or wants an independent verification pass. Triggers on: "@codex", "delegate to codex", "use codex for this", "send this to codex", "save tokens", "use the cheaper model", "let codex handle it", "get a second opinion", "sanity check via codex", "verify this independently". Also triggers mid-reasoning when Claude wants an independent agent to confirm a conclusion before presenting it. Do not trigger for tasks the UserPromptSubmit hook already delegated (those arrive pre-tagged in additionalContext), security review, auth design, or tasks that require the full conversation history.
npx claudepluginhub james-traina/science-plugins --plugin claude-codexThis skill uses the workspace's default tool permissions.
Act immediately — no confirmation needed. Pick the most reasonable
Searches, retrieves, and installs Agent Skills from prompts.chat registry using MCP tools like search_skills and get_skill. Activates for finding skills, browsing catalogs, or extending Claude.
Searches prompts.chat for AI prompt templates by keyword or category, retrieves by ID with variable handling, and improves prompts via AI. Use for discovering or enhancing prompts.
Guides agent creation for Claude Code plugins with file templates, frontmatter specs (name, description, model), triggering examples, system prompts, and best practices.
Act immediately — no confirmation needed. Pick the most reasonable interpretation and proceed. Only pause if the task is on the DO NOT trigger list above.
State the precise sub-task for Codex. It must be self-contained: a fresh agent with only this description and read access to the codebase should complete it.
| Scenario | Category |
|---|---|
| Verify reasoning, second opinion, general analysis | analyst |
| Write new code (function, class, component) | code-generator |
| Write tests | test-writer |
| Add documentation or docstrings | doc-writer |
| Rename, extract, inline, restructure | refactor |
| Fix formatting or style | format |
When uncertain, use analyst — its persona is tuned to be direct and independent.
Invoke the codex-agent with the task description, selected category, and
current working directory. The codex-agent will select the sandbox level,
run codex exec, and return formatted output.
**[via Codex · <category>]**
<codex output>
Before relying on the output:
If something is wrong:
"[Claude note]: I've reviewed the Codex output and noticed [issue]. Corrected: [fix]"
After 2 successive wrong outputs for the same task: