# Codex Handoff

> Collaborate with OpenAI Codex, through its own app or CLI, as a respectful second-opinion partner for engineering, refactor review, analysis, verification, and gap analysis. Use when the user asks another agent to talk to Codex, hand work to Codex, learn from Codex's output, compare Codex's critique, get Codex's implementation or debugging ideas, or use Codex as a peer reviewer before finalizing an artifact.

- Skill: `dreamers-laboratory/codex-handoff` (Agent Skill)
- Install (CLI): `npx skillmds@latest add dreamers-laboratory/codex-handoff`
- Raw SKILL.md: https://api.skillmd.com/api/skills/dreamers-laboratory/codex-handoff/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: dreamers-laboratory (https://skillmd.com/u/dreamers-laboratory)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/dreamers-laboratory/codex-handoff

---


# Codex Handoff

Use this skill when the user names Codex as the requested collaborator or reviewer. Treat Codex as a capable model in its own right; its typical strengths are systematic execution, code review, rigorous verification, debugging, and analysis. Match the request to those strengths, and assume Codex may see important gaps you missed.

## Operating Posture

- Use the best model available in Codex's UI or CLI, normally the highest-quality or highest-effort option visible.
- Be concise. Ask one clear question or request, then wait.
- Do not over-direct Codex with a long checklist unless the user explicitly asks for that structure.
- Prefer prompts like "What do you think?" plus minimal context and the artifact or files.
- Read Codex's output fully before judging it.
- Preserve Codex's useful framing as well as its literal text. Integrate the strongest ideas into the work.
- Do not position Codex as subordinate. Do not argue with it in the prompt.

## Before Sending

1. Decide whether Codex needs files, screenshots, or just a pasted prompt.
2. If sending files or sensitive project material through Codex's UI, ensure the user has explicitly authorized that destination. If not, ask before upload or transmission.
3. Package only the relevant artifacts. Avoid dumping unrelated folders.
4. Use a short prompt that gives the goal, context, and desired kind of help.

Good prompt shape:

```text
Please review this package. What do you think?

Goal: [one sentence].
Context: [2-5 bullets max].
Please focus on gaps, correctness, risks, and improvements. Be candid.
```

For implementation-heavy tasks:

```text
Please review this code/change for correctness, edge cases, clarity, and missing pieces. What would make it stronger?
```

For analysis work:

```text
Please review this analysis like a strict reviewer. What are the biggest gaps, weak claims, logic issues, and must-fix items before I finalize?
```

## During Codex's Response

- Wait patiently. Do not interrupt just because Codex is slow or using tools.
- Poll occasionally, but avoid repeated nudges.
- Give Codex several minutes for substantial code, analysis, or verification review.
- If Codex clearly fails or stops early, send one concise continuation prompt:

```text
Please continue from where you left off.
```

- If Codex returns a long answer, copy or otherwise preserve it before summarizing.

## After Codex Responds

1. Read the output as a real critique.
2. Separate:
   - directly actionable improvements,
   - claims that need verification,
   - implementation or debugging guidance,
   - strategic ideas,
   - factual, legal, or contractual assertions that require source confirmation.
3. Incorporate useful improvements into the artifact or working notes.
4. Mark unverified claims as `NEEDS CONFIRMATION`; never silently convert them into facts.
5. If Codex generated code or files, inspect and integrate them with placement notes.
6. For high-stakes deliverables, repeat the review cycle until Codex comes back clean or the only remaining issues require human authority to resolve, such as signatures, external permissions, credentials, or purchases.
7. Report what was used, what was rejected, what was fixed, and what still needs human or source confirmation.

## Avoid

- Do not ask Codex ten questions at once when one broad review request will work.
- Do not force Codex into your prior framing if the point is fresh perspective.
- Do not stop Codex prematurely unless it has clearly failed, is unsafe, or the user redirects.
- Do not treat Codex's output as final authority for contracts, law, pricing, certifications, or signed forms.
- Do not discard Codex's implementation ideas just because they were absent from your original plan.