| name | delegate |
| description | 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. |
Delegation Assessment & Dispatch
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.
1. Identify What to Delegate
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.
- Verification request → frame as: "Independently verify: [your conclusion].
Return 'Confirmed: ' or 'Correction: <what's right>'."
- Budget offload → frame as you would frame it to any capable engineer.
2. Select Expert Category
| 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.
3. Dispatch to codex-agent
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.
4. Present Results
**[via Codex · <category>]**
<codex output>
- Verification result: present as "Codex confirms: …" or "Codex disagrees: …"
and act accordingly. Never discard a correction silently.
- Never present Codex output as your own work.
5. Validate
Before relying on the output:
- Code: check for syntax errors, missing imports, obvious logic flaws.
- Analysis: check that conclusions follow from the evidence.
- Verification: check that confirmation/correction is internally consistent.
If something is wrong:
"[Claude note]: I've reviewed the Codex output and noticed [issue]. Corrected: [fix]"
Escalation
After 2 successive wrong outputs for the same task:
- Stop delegating for this task.
- Handle it directly.
- Tell the user: "Codex produced inconsistent results here. Handling it directly."
