Standing Mandates
- NEVER put your own answer, hypothesis or ranking in the brief. A second opinion that reads your conclusion agrees with it; the point is a reader who has not seen it.
- ALWAYS hand over the question and the material (paths to read, the diff, the error), the constraints that are facts, and the shape of answer wanted — nothing else.
- ALWAYS run it read only through the script unless the user asked for Claude to do the work: Claude gets Read, Grep and Glob, every other tool is denied, and no session is kept.
--actruns it with--permission-mode bypassPermissionsinstead: every tool, no prompt, in--cwd(user settings and their hooks stay loaded). Use it only on the user's ask, in a tree whose other changes are committed or named, never for a second opinion. - ALWAYS check an
--actrun yourself before calling it done: thechangedlist and its diff,head_moved, and the task's own check run fresh. Claude's "done" is a claim, not evidence. - NEVER adopt the other opinion because it disagrees, nor dismiss it because it does. Each disagreement is settled by evidence you check (a file, a test, a command), and the report says which.
- NEVER ask again from inside a second opinion: the script refuses a nested call (
SECOND_OPINION_DEPTH). Claude's read-only child has no shell, so it cannot start the other CLI by hand; an--actchild has one, so its brief says not to call another agent.
Use Claude — a second opinion from the other model
Codex (or any session with a shell) asks Claude the same question, cold, and puts the two answers side by side.
Process
- Frame the question. One question the user cares about, in their words where possible. Write it before looking at your own answer again.
- Write the brief to a scratch file (never into the repo):
- the question;
- what to read: paths, a diff (
git diff …), the error text, the doc; - facts that constrain the answer (versions, what cannot change);
- the answer shape: verdict first, reasons with the file:line they rest on, what would change it. Leave out: your answer, your suspect, "we think", leading phrasing ("isn't it true that …").
- Run it from the directory the material lives in:
Claude has only Read, Grep and Glob (node <this skill's dir>/../../scripts/ask.mjs --to claude --prompt-file <brief.md> --cwd <dir> [--act] [--attach a,b] [--model opus|sonnet] [--timeout 600]--tools), no MCP server, no session kept;--attachputs the files' text into the brief when Claude should not open anything else. One JSON line back:answer, orerror(not installed, not signed in, timeout, nested call). An error is reported as it is; the work goes on without the second opinion and says so. - Compare. Your answer and Claude's, point by point: agree · disagree · only one of us raised. For each disagreement, check the evidence yourself and settle it, or mark it open.
- Report with the template below; then act only on what survived step 4.
When Claude does the task (--act)
- Brief = the task, the files it may change, the check that proves it (
node --test …), and "do not commit, do not push, do not call another agent". Here your own plan may go in: it is an order, not a second opinion. - Run with
--act(no--attach: it has every tool). - Check
changed(git status lines that appeared or changed;nulloutside git) andhead_moved. Read the diff of each file, rerun the check, and revert what is outside the brief. - Report what Claude changed, what you kept or reverted, and the check's count.
Output Template
## Second opinion — claude (<model>, <seconds>s)
Question: <one line>
| Point | Codex | Claude | Settled by | Verdict |
|---|---|---|---|---|
| … | … | … | <file:line / test / command> | agree · Codex · Claude · open |
Only Claude raised: <items, each checked or marked unchecked>
What changes: <the decision or the fix, or "nothing — both agree on …">
What Claude Does / What You Do
| Claude | You |
|---|---|
| The asking agent (Codex here) writes a brief without its own answer, runs Claude read only, compares point by point | Say what the question is when it is not obvious |
| Settles each disagreement with evidence, says which side held | Decide the open ones |
With --act, checks Claude's changes and the task's test itself, reverts what is outside the brief |
Ask for it ("claude한테 시켜"); keep unrelated work committed |
| Reports a failed call plainly and continues without it | Install and sign in to claude once |
Related Skills
agents:use-codex— the same, from Claude Code to Codexagents:dispatching-parallel-agents— several independent jobs at once, inside one modelthink:devils-advocate— an attack on a plan from the same model; it calls this automatically for a cold objection setthink:brainstorming— its debate mode uses the other model as a debater, automatically