4
steps per round: 3
max iterations: 1
SKILL.md file

Loop kaise chalta hai

The loop

Claude: PLAN  -->  Codex: EXECUTE  -->  Claude: CHECK  -->  Codex: IMPROVE
                                              |                    |
                                  sab pass? STOP  <----- nahi to wapas CHECK

Role	Kaun	Kaam
Planner	Claude	Task ko chhote steps + "done" criteria mein todta hai
Executor	Codex	Plan ke mutabiq code likhta hai
Checker	Claude	Diff padhta hai, tests chalata hai, issues ki list banata hai
Improver	Codex	Sirf us list ke issues fix karta hai

Stop condition: checker kahe "sab criteria pass" (tests green + koi issue nahi).
Max iterations: 3 (cap zaroor lagao, warna loop chalta reh sakta hai aur usage jalta hai).
Roles ulta bhi kar sakte ho (Codex plan, Claude execute). Concept same hai: ek banata hai, doosra parakhta hai.

Ye cheezein ready rakho

Claude Code account: Pro, Max, Team, Enterprise ya Console. Free claude.ai plan mein Claude Code nahi hai (docs ke mutabiq).
Codex ke liye ChatGPT account se sign-in (ya API key). Kaunsa plan Codex deta hai aur kitni limit: check current limits OpenAI ke pricing page pe.
Node.js 22+ (npm install ke liye) aur Git.
Ek project folder jo git repo ho.

Claude Code install

Windows PowerShell

irm https://claude.ai/install.ps1 | iex

macOS / Linux / WSL

curl -fsSL https://claude.ai/install.sh | bash

Naya terminal kholo, claude --version chalao, phir claude se login karo.

Codex CLI install

Terminal

npm install -g @openai/codex
codex --version
codex        # pehli baar: "Sign in with ChatGPT"

macOS/Linux pe OpenAI ka installer bhi hai: curl -fsSL https://chatgpt.com/codex/install.sh | sh.

Claude Codex ko kaise bulata hai

Claude Code Bash se koi bhi CLI chala sakta hai. Codex ka non-interactive mode codex exec hai:

Claude Code chalata hai

cat .loop/task.md | codex exec --sandbox workspace-write -

--sandbox workspace-write = Codex sirf project ki files edit kar sake. Default read-only hai.
- ka matlab prompt stdin se aaye (quotes ki jhanjhat nahi).
danger-full-access kabhi mat use karo.

Skill folder banao aur SKILL.md paste karo

Terminal

mkdir -p ~/.claude/skills/claude-codex-loop

(Windows: %USERPROFILE%\.claude\skills\claude-codex-loop.) Sirf ek project ke liye: .claude/skills/claude-codex-loop/ project ke andar. Phir usme SKILL.md banao:

SKILL.md

---
name: claude-codex-loop
description: Run a plan/execute/check/improve loop where Claude plans and checks and Codex executes and improves. Use when the user asks for the Claude x Codex loop.
disable-model-invocation: true
argument-hint: "[task description]"
allowed-tools: Bash(codex exec *) Bash(git *)
---

Task: $ARGUMENTS

Rules: max 3 iterations. Never touch .env or secrets. Codex runs only with --sandbox workspace-write.

1. PLAN (you, Claude): write .loop/plan.md with small steps and a "Done when" list
   of testable criteria (tests pass, lint clean, specific behavior works).
   Make sure we are on a new git branch first.
2. EXECUTE (Codex): run
   cat .loop/plan.md | codex exec --sandbox workspace-write -
   with the instruction "Implement this plan. Do not change anything outside it."
3. CHECK (you, Claude): read `git diff`, run the project's tests/lint, compare with
   "Done when". Write .loop/issues.md: PASS or a numbered list of concrete problems.
4. IMPROVE (Codex): if issues exist, run
   cat .loop/issues.md | codex exec --sandbox workspace-write -
   with "Fix only these issues." Then go back to step 3.
5. STOP when CHECK says PASS or after 3 iterations. Report: iterations used, what
   changed, anything still failing. Do not commit; leave the diff for the human.

Claude Code skills docs ke mutabiq disable-model-invocation: true ka matlab skill sirf tum manually chalate ho. allowed-tools sirf us ek turn ke liye pre-approve karta hai.

Chalao

Claude Code kholo (project folder mein) aur:

Claude Code

/claude-codex-loop add a /health endpoint with a test

Worked example

Task: "Express app mein /health endpoint + test."

Plan: Claude .loop/plan.md banata hai: route add, Jest test, "Done when: npm test pass, GET /health = 200 {status:"ok"}".
Execute: Codex route aur test likhta hai.
Check: Claude tests chalata hai. Test fail: response JSON key galat. issues.md mein 1 issue.
Improve: Codex sirf wo fix karta hai.
Check (round 2): PASS. Loop stop. Tum git diff khud review karte ho.

(Ye illustration hai, koi guaranteed result nahi.)

Safety

Diff khud review karo commit se pehle. Checker bhi AI hai, galat ho sakta hai.
Prod secrets mat do: .env, API keys, production DB URL agents ke paas na hon. Alag dev keys use karo.
Iteration cap (3) aur chhota task: loop ka har round dono taraf usage khata hai.
Naye git branch pe chalao taake wapas ja sako.
danger-full-access ya "sab approve" mode avoid karo.

Common Mistakes

Vague task: "app better banao". "Done when" measurable likho.
Cap nahi lagaya: loop atak ke usage jala deta hai.
Checker ke paas tests hi nahi: phir "PASS" ka matlab kuch nahi.
Executor ko poora folder + secrets access de diya.
Ek round mein bohot bada task. Chhote tasks mein todo.
Codex login/install bhoole, codex: command not found: codex --version pehle check karo.
Diff review kiye bina merge.

Checklist

Claude Code install + login (claude --version)
Codex CLI install + sign-in (codex --version)
Project git repo hai, naya branch bana
~/.claude/skills/claude-codex-loop/SKILL.md paste kiya
.env/secrets agents se door
Max iterations = 3
Task chhota, "Done when" measurable
/claude-codex-loop <task> chalaya, diff review kiya

Using Claude Code or Codex?

Paste this. Let your agent set it up.

Prompt

Set up the Claude x Codex loop skill from https://theaikhan.com/guides/claude-codex-loop
1. Run `claude --version` and `codex --version`. If one is missing, stop and tell me how to install it.
2. Create ~/.claude/skills/claude-codex-loop/SKILL.md with the exact SKILL.md from the guide.
3. Never read or touch .env files or secrets.
4. When done, tell me to open my project on a new git branch and run:
   /claude-codex-loop <small task with a clear "done when">