codex

name	codex
description	Send a file to codex (GPT-5.5) for review. Reports feedback — no changes until the user approves.
argument-hint	["file_path(s) or leave blank for in-context content"]
allowed-tools	Read, Bash, Glob, AskUserQuestion

Codex Review

Send content to codex for a second opinion. Report the feedback verbatim. Don't change anything until the user approves.

Process

1. Resolve the content. Three modes:

   • File path(s) given: Read and concatenate them.
   • Content in conversation context: The user has been working on something in-context (e.g., a draft, a plan, a code block). Gather it.
   • Nothing clear: Ask the user what to send.

2. Build the prompt. Compose a single prompt that includes:

   • A direct review request tailored to the content type (prose, code, plan, etc.)
   • All the content, clearly delimited.
   • For prose meant for readers outside the author's bubble, a legibility lens: ask codex to read as a smart non-specialist and name the first place they'd bounce, and whether they can tell what the thing is and does. The author and an in-context assistant are both too deep to see this (the curse of knowledge); an external model isn't.

3. Send to codex. Pipe the prompt via stdin to avoid shell quoting issues:

cat <<'PROMPT_EOF' | codex exec -
Review the following. What works, what doesn't, what to cut, what to strengthen. Be direct — no preamble, no praise sandwich.

---

<content here>
PROMPT_EOF

For multiple files, include each with a filename header:

cat <<'PROMPT_EOF' | codex exec -
Review the following files together. What works, what doesn't, what to cut, what to strengthen. Be direct — no preamble, no praise sandwich.

--- file: path/to/first.md ---
<contents>

--- file: path/to/second.md ---
<contents>
PROMPT_EOF

4. Report the feedback to the user exactly as codex returns it. Don't summarize, reframe, or editorialize.
5. Wait for the user to say what to act on. Don't apply changes unprompted.

Rules

• Don't roleplay in the prompt. No "You are a senior editor." Just ask directly.
• Don't add your own review on top. The point is a second opinion, not two opinions stacked.
• If codex flags something with low confidence, research it before dismissing. Structural soundness is not a substitute for evidence.
• Two passes max. If the second pass has no structural notes, the content is ready.
• Always pipe via stdin (| codex exec -) rather than inline quoting, to avoid shell escaping issues with large content.