# Codex

> Guides operator-owned delegation to Codex when the task fits a bounded coding lane, review pass, or worktree-isolated implementation track.

- Skill: `agentic-in/codex` (Agent Skill)
- Install (CLI): `npx skillmds@latest add agentic-in/codex`
- Raw SKILL.md: https://api.skillmd.com/api/skills/agentic-in/codex/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: agentic-in (https://skillmd.com/u/agentic-in)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/agentic-in/codex

---


# Codex

Use this built-in skill when the user wants a task executed through Codex or when a dedicated coding lane is the cleanest way to isolate work.

## Core rules

- Prefer repo-native instructions and validation over generic agent folklore.
- Delegate only work with a clear boundary, write scope, and success condition.
- Use worktrees or isolated branches when the task is not safe inside the current checkout.
- Treat Codex output as a candidate change set that still needs human review and repo validation.

## Default workflow

1. Decide whether the next blocking step should stay local or move to Codex.
2. Define the task packet: scope, files, tests, branch or worktree, and ship expectations.
3. Review the returned changes against the harness and any user-owned edits.
4. Validate, integrate, and summarize the outcome in operator language.

## Guardrails

- Do not delegate vague exploratory work just to create agent theater.
- Do not send Codex into a dirty shared checkout without a coordination plan.
- Do not skip validation because the delegated run looked confident.