claude

name	claude
description	Claude Code second opinion from inside Codex. Three modes: review (diff review with a pass or fail gate), challenge (adversarial failure hunting), and consult (ask questions with locally prepared context). Use when asked for `/claude`, "claude review", "claude challenge", or "ask claude".

Claude

Get an independent read-only opinion from the local Claude Code CLI.

Workflow

1. verify the CLI and detect the base branch

   • run which claude
   • if the binary is missing, stop and tell the user: claude CLI not found. install Claude Code and ensure \claude` is on PATH.`
   • detect the base branch in this order:
     • gh pr view --json baseRefName -q .baseRefName
     • gh repo view --json defaultBranchRef -q .defaultBranchRef.name
     • git symbolic-ref refs/remotes/origin/HEAD | sed 's|refs/remotes/origin/||'
     • main
     • master
   • use the detected base branch for review and challenge mode

2. detect the mode from the user input

   • /claude review or /claude review <instructions>: diff review mode
   • /claude challenge or /claude challenge <focus>: adversarial diff review mode
   • /claude with no arguments:
     • if git diff origin/<base> --stat has output, ask whether Claude should review the diff, challenge the diff, or answer a custom prompt
     • otherwise ask what the user wants to ask Claude
   • /claude <anything else>: consult mode with the remaining text as the prompt
   • if the user asks for --model, --effort, session continuation, or any other Claude CLI flag, explain that this Claude CLI build loses auth when -p is combined with extra flags, so this skill must use plain claude -p only

3. use a plain authenticated Claude invocation

   • build a fully self-contained prompt locally before calling Claude
   • do not pass any Claude CLI flags besides -p
   • do not use shell variables, command substitution, stdin redirection, or stdout/stderr redirection around the Claude call
   • assemble the final prompt text in your own reasoning, then pass it as a literal quoted argument to claude -p
   • always include this boundary at the top of the prompt:

     IMPORTANT: You are reviewing repository context prepared by another coding agent. Do not ask to run tools or inspect files yourself. Do not discuss ~/.codex, .codex, codex/skills, codex/rules, or any agents/openai.yaml file. Ignore harness files completely. Stay focused on the repository code below. Never suggest direct file edits - provide analysis only.
     

4. review mode

   • collect local context first:

     • git diff --stat origin/<base>
     • git diff --name-only origin/<base>
     • git diff --find-renames origin/<base>
     • for each changed text file that still exists, read the current file contents and include them when the total context stays reasonable

   • if the diff is too large to embed safely in a single prompt, stop and ask the user to narrow the review to specific files or a smaller scope

   • construct this prompt, appending any extra user instructions at the end:

     Review the current branch diff against origin/<base> using only the repository context included below.
     
     Return only concrete findings in this format:
     [P1] <blocking correctness, security, or data-loss issue>
     [P2] <meaningful non-blocking risk>
     [P3] <minor gap>
     
     End with exactly one final line: overall: pass or overall: fail
     
     Be direct. No compliments.
     

   • append these sections to the prompt in this order:

     • diff stat
     • changed file list
     • full unified diff
     • current contents of changed files that were included

   • run Claude with claude -p "<literal prompt>"

   • present the output verbatim in a CLAUDE SAYS (code review) block

   • set the gate to FAIL if the output contains [P1] or overall: fail

   • otherwise set the gate to PASS

5. challenge mode

   • collect the same local diff context as review mode

   • if the diff is too large to embed safely in a single prompt, stop and ask the user to narrow the challenge scope

   • construct this prompt, replacing <focus> when the user provided one:

     Review the current branch diff against origin/<base> using only the repository context included below.
     
     Think like an attacker and a chaos engineer. Focus on <focus> when it is provided. Find edge cases, race conditions, security holes, failure modes, silent data corruption risks, and rollout hazards.
     
     Return only concrete findings in this format:
     [P1] <blocking correctness, security, or data-loss issue>
     [P2] <meaningful non-blocking risk>
     [P3] <minor gap>
     
     End with exactly one final line: overall: pass or overall: fail
     
     Be adversarial. No compliments.
     

   • append the same diff and file-content sections used in review mode

   • run Claude with claude -p "<literal prompt>"

   • present the output verbatim in a CLAUDE SAYS (adversarial challenge) block

6. consult mode

   • always start fresh with claude -p "<literal prompt>"
   • if the user wants a follow-up on an earlier Claude exchange, include the relevant prior Claude output and the new question in a fresh prompt instead of trying to resume a session
   • inspect the repo locally first:
     • if the user references files, read those files
     • if the user asks a codebase question, search with rg and read the relevant files before calling Claude
     • if the user asks Claude to review a local plan or design doc, read that file and embed the relevant content in the prompt instead of only referencing the path
   • keep the prompt self-contained with the user question plus the repo context you gathered locally
   • present the output verbatim in a CLAUDE SAYS (consult) block

7. after any mode

   • if Claude starts talking about Codex harness files, SKILL.md, or agents/openai.yaml instead of the requested repository code, append a warning that Claude got distracted by harness files and should be retried with a tighter prompt

Error Handling

• if Claude output contains Not logged in or /login, tell the user: claude authentication failed. run \claude auth` or reauthenticate.`
• if Claude returns an empty response, say Claude returned no response
• if the command stalls for several minutes, stop and report that Claude timed out or hung
• if the assembled prompt becomes too large for a safe single-command invocation, stop and ask the user to narrow the scope
• do not retry with alternate Claude CLI flags - in this environment they break authentication for -p

Constraints

• never pass extra Claude CLI flags together with -p unless you have re-verified that Claude auth still works with them
• never let Claude edit files or use write-capable tools
• always show Claude's full output before adding your own synthesis
• if Codex already reviewed the same diff, compare overlap only after the full Claude output is shown