Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Ask OpenAI Codex (GPT-5) a one-off question about the current repo and get its answer back, without spinning up a subagent. Use when the user says "ask codex", "what does GPT think", "get a second opinion on this", or when you want an independent read on a design call, a tricky bug, or an unfamiliar area of the codebase.
.claude/skills/sateezg-ask-codex-a-question/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -31% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -36% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -35% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -54% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 61% | 0% |
For a single question with a single answer. Anything multi-step or long-running belongs to a subagent (codex-second-opinion, codex-reviewer, codex-debugger).
bashcodex-run -C <repo-dir> "<self-contained question>"
codex-run prints only Codex's final answer on stdout (it captures the final message via codex exec -o rather than dumping the whole event log). It defaults to a read-only sandbox, so Codex can read the repo but cannot change anything.
Codex has zero knowledge of this conversation. Every question must stand alone:
line to look at", "answer yes/no then justify in three sentences").
bashcodex-run -C /path/to/repo \ "In src/auth/session.ts, refreshToken() occasionally returns a token that fails validation immediately. I suspect a clock-skew issue in the exp comparison. Read that file plus src/auth/jwt.ts and tell me the single most likely root cause and the exact line. Be concise."
-s workspace-write let Codex modify files (only when the user asked for that)
-m gpt-5-codex model override
-r continue the previous codex session (follow-up question)
--timeout 1800 raise the 900s default for a big repo sweep
--raw also print Codex's full event log to stderr, for debuggingUse -r for a genuine follow-up — it keeps Codex's own context and avoids paying to re-read the repo:
bashcodex-run -C /path/to/repo -r "Now show me the minimal patch for that."
your own take separately if you disagree — the value here is having two independent opinions, and blending them destroys that.
Verify any specific claim against the repo before acting on it.
-s workspace-write on your owninitiative.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→pass | 15,303 | 13,544 | -11% | 1 | 1 | 0% | 1,467 | 1,011 | -31% | 0 | 0 | — |
case-03 | fail→pass | 12,841 | 4,605 | -64% | 1 | 1 | 0% | 2,201 | 1,398 | -36% | 0 | 0 | — |
case-01 | fail→pass | 15,499 | 7,831 | -49% | 1 | 1 | 0% | 1,697 | 1,099 | -35% | 0 | 0 | — |
case-04 | fail→fail | 16,766 | 15,529 | -7% | 1 | 1 | 0% | 1,352 | 1,119 | -17% | 0 | 0 | — |
case-05 | fail→fail | 3,999 | 13,668 | +242% | 1 | 1 | 0% | 702 | 1,142 | +63% | 0 | 0 | — |
case-06 | fail→fail | 7,393 | 17,752 | +140% | 1 | 1 | 0% | 662 | 1,913 | +189% | 0 | 0 | — |
case-07 | fail→pass | 21,890 | 9,420 | -57% | 1 | 1 | 0% | 2,694 | 1,243 | -54% | 0 | 0 | — |
case-08 | fail→fail | 15,932 | 13,993 | -12% | 1 | 1 | 0% | 290 | 1,102 | +280% | 0 | 0 | — |
case-09 | fail→fail | 18,266 | 18,890 | +3% | 1 | 1 | 0% | 2,360 | 1,136 | -52% | 0 | 0 | — |
case-10 | fail→pass | 9,525 | 3,033 | -68% | 1 | 1 | 0% | 692 | 1,113 | +61% | 0 | 0 | — |
case-11 | pass→pass | 13,444 | 7,152 | -47% | 1 | 1 | 0% | 1,310 | 843 | -36% | 0 | 0 | — |
case-12 | fail→pass | 8,136 | 4,080 | -50% | 1 | 1 | 0% | 1,206 | 1,315 | +9% | 0 | 0 | — |
case-13 | pass→pass | 9,306 | 8,834 | -5% | 1 | 1 | 0% | 603 | 1,143 | +90% | 0 | 0 | — |
case-14 | pass→pass | 3,497 | 3,334 | -5% | 1 | 1 | 0% | 641 | 1,240 | +93% | 0 | 0 | — |
case-15 | fail→pass | 11,747 | 7,598 | -35% | 1 | 1 | 0% | 1,100 | 1,042 | -5% | 0 | 0 | — |
case-16 | pass→pass | 7,354 | 9,341 | +27% | 1 | 1 | 0% | 1,406 | 1,294 | -8% | 0 | 0 | — |
case-17 | pass→pass | 12,310 | 9,755 | -21% | 1 | 1 | 0% | 1,035 | 1,513 | +46% | 0 | 0 | — |
case-18 | pass→pass | 10,668 | 8,757 | -18% | 1 | 1 | 0% | 864 | 1,122 | +30% | 0 | 0 | — |
case-19 | fail→pass | 9,071 | 3,835 | -58% | 1 | 1 | 0% | 1,531 | 1,337 | -13% | 0 | 0 | — |
case-20 | pass→pass | 19,265 | 9,677 | -50% | 1 | 1 | 0% | 2,061 | 1,180 | -43% | 0 | 0 | — |
case-21 | fail→fail | 13,099 | 15,649 | +19% | 1 | 1 | 0% | 310 | 1,198 | +286% | 0 | 0 | — |
case-22 | fail→pass | 10,134 | 8,884 | -12% | 1 | 1 | 0% | 915 | 1,157 | +26% | 0 | 0 | — |
case-23 | pass→pass | 19,221 | 8,090 | -58% | 1 | 1 | 0% | 2,175 | 1,090 | -50% | 0 | 0 | — |
case-24 | fail→pass | 7,061 | 7,721 | +9% | 1 | 1 | 0% | 1,035 | 980 | -5% | 0 | 0 | — |
case-25 | fail→pass | 11,475 | 6,940 | -40% | 1 | 1 | 0% | 1,099 | 951 | -13% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 25 cases were attempted, and 20 counted toward the lift figure. The other 5 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +44 percentage points is the difference between those two pass rates over the 20 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.