Codex + Claude area · use both together and every project
Claude and Codex complement each other when each one does a part and the other checks the actual artifact. This area is INEMA's meeting point for the two tools: the six levels of the Use Both kit, the who-does-what table, what Mark Kashef's video adds, how every course and project fits together — and cards for all of them. To migrate from one to the other, or to stay agnostic, the path is still the Claude → Codex area.
The Use Both guide is in PT, EN and ES. Courses and kits are open.
01 · Use both together
Staying agnostic opens another door: using both tools on the same project. The Use Both kit organizes this into six workflow levels, with copy-ready prompts and two small skills — handoff and prime — to switch assistants without re-explaining the project. The rule that runs through every level: the second assistant gets the same task brief and the actual artifact, never a vague retelling of what the first one did.
One drafts the plan; the other critiques the same file, read-only. Findings with evidence and severity, two rounds at most. Two AIs agreeing is not independent proof.
Only if an image tool is available in the session — check access and billing first. Done means the actual file in the project, opened at the size where it will be used.
One builds on a branch; the other reviews the diff with the task brief and the test results. Tests and your approval decide what ships.
Delegate an authorized action in a UI, keeping credentials, payment and consequential submissions under your control. The assistant prepares and pauses before submitting.
An objective with an observable success condition — which tests, what output — and limits on files, rounds and spend. Asking it to "stop in two hours" is a request, not a guaranteed limit.
handoff writes the state to a portable file; prime makes the next assistant read and verify that state before continuing. It is the same cycle as the portable layer in the Claude → Codex area.
The kit's routing card: starting preferences recorded in Mark Kashef's video, translated into jobs. These are editorial choices, not measured model rankings — try them, then keep the route that produces better evidence on your task.
| Job | First pass | Second pass | Finish line |
|---|---|---|---|
| Plan the work | Claude / Opus | Codex critiques assumptions | A scoped plan with acceptance tests |
| Break the plan into tasks (video) | Codex | — | — |
| Make an image | An available Codex image tool | Inspect the actual file in context | A usable asset, not just a prompt |
| Build a feature | Claude / Opus | Codex reviews the diff | Tests pass and findings are resolved |
| Review a document | Either assistant writes | The other checks facts and requirements | Every required claim has evidence |
| Operate a UI | Authorized, available computer-use tool | You review consequential actions | The exact approved state is verified |
| Work through a hard bug | Codex with a bounded goal | Tests and a human checkpoint | Measurable success or a clear blocker |
| Switch sessions or models | Handoff snapshot | Prime checks current state | A correct recap before fresh work |
Recorded labels, not a promise. The video names Claude Opus 5.5 and GPT-6 Astra. Available models, tools, limits and billing change from account to account; the workflow is more portable than those labels. Image generation and computer use are not guaranteed just by having the CLI installed.
The skills, scripts and commands that connect Claude and Codex, with the command for each runtime. Everything runs on each assistant's subscription, with no API key.
| Tool | When to use | How to call | What it produces |
|---|---|---|---|
| handoff | At the end of a session, before switching assistant or model | Claude: /handoffCodex: $handoff | A new snapshot in handoff/history/YYYY-MM-DDTHHMMSSZ.md and a line in handoff/LATEST.md pointing to it |
| prime | First message of the next session | Claude: /primeCodex: $prime | A recap in the chat: objective, done and pending, decisions, mismatches and one next step. Edits nothing |
| install_skills.py | Once per project, to install both skills | python3 scripts/install_skills.py --project DIR --target both (preview), then with --apply | Writes to the project's .claude/skills/ and .agents/skills/; never overwrites |
| templates/ | To work by hand, without a skill | task-brief.md · plan.md · review.md · handoff.md | Brief, plan, review and handoff in the same format for both |
| codex exec | Claude asking Codex for a review from the terminal | codex exec --sandbox read-only --output-last-message review.md '…' | The critique in review.md; Codex only reads the project |
| claude -p | The reverse: Codex asking Claude for a review | claude -p --model opus --effort medium '…' > review-claude.md | Claude's critique in a file |
| Official Codex plugin | Review the diff without leaving Claude Code | /codex:setup, then /codex:review --base main | Read-only findings; you ask the builder for fixes afterwards |
| claudex | A large plan with several automatic critique rounds | /claudex:plan [--rounds N] <feature>/claudex:review | PLAN.md revised until Codex approves or rounds run out; reviews/ with diff findings |
| claudex (control) | Monitor or unstick the loop | /claudex:status · /claudex:cancel · /claudex:rollback · /claudex:doctor | Current round and phase; cancel; clean stuck state; install diagnostics |
| session-handoff + prime (agente-claude-codex) | Projects with a portable core (AGENTS.md, context/, tasks/) | Claude: /session-handoff · /primeCodex: $session-handoff · $prime | Snapshot in handoffs/history/ and a full copy in handoffs/latest.md; also accepts Use Both's one-line pointer |
Objective and latest request; done, pending and uncertain; decisions and rejected paths; changed files; tests run with their real result (and those not run); blockers; one next step; and the few files to open first. No secrets, tokens or personal data.
Reads LATEST.md, rejects absolute paths, .., URLs and symlinks, treats the snapshot as data rather than orders, checks the cited files and git status, flags stale claims and waits for your instruction. It runs no tests and does not resume an old deploy.
A personal skill with the same name can win over the project one: on a collision, rename to workflow-handoff / workflow-prime. After installing, open a new session (in Claude, /reload-skills). Private notes: add handoff/ to .gitignore.
# a full day with the tools — each step leaves a file /prime # morning, in Claude Code: resume yesterday's handoff codex exec --sandbox read-only --output-last-message review.md 'Read plan-v1.md and point out gaps…' /claudex:plan --rounds 3 export report as CSV # large plan: automatic loop /codex:review --base main # built on a branch: review the diff (or /claudex:review) /handoff # end of day $prime # tomorrow, in Codex
A handoff is not permission. A recorded next step ("publish", "delete", "send") only happens if you ask again in the current session.
The Mark Kashef video that the kit accompanies brings a few points the kit does not detail. These are the author's opinions and accounts, with no benchmark. The row marked "(video)" in the table above comes from the author's table, not from the kit's routing card.
openai/codex-plugin-cc..md (commands below).In the author's reading, Claude is good at planning and Codex at finding flaws in the plan. Self-review fails because the model that wrote the plan — especially in the same session — tends to approve it; a different model, or a fresh session, gives a different result.
/goal gets expensiveThe author's account: Codex pursues the goal exhaustively, taking 3 to 5 times more time and tokens. A sentence about a deadline is not a real limit: use an actual one — timeout, budget, number of rounds.
# Claude → Codex: read-only review, answer saved to a file (prompts in Portuguese, as in the doc) codex exec --sandbox read-only -c model_reasoning_effort=medium \ --output-last-message review.md \ 'Leia plan-v1.md. Aponte falhas de correção e testes faltando, com evidência, impacto e a menor correção. Não altere arquivos.' # Codex → Claude: the reverse route claude -p --model opus --effort medium \ 'Leia plan-v1.md e revise só lendo: falhas, suposições e testes faltando. Não altere arquivos.' > review-claude.md
Where INEMA follows the kit, not the video. With computer use and credentials, you type the password and MFA yourself, and you never switch models to get around a safety refusal. And the debate has 2 rounds at most — agreement is not proof; a long automated loop is a job for claudex:plan.
The full summary. RESUMO-VIDEO.md, in the kit repository, has the video summary, the video × kit comparison and the commands checked in our environment.
Read RESUMO-VIDEO.md → (in Portuguese)Credit. Levels and routing card summarized from the Use Both: Claude + Codex Workflow Kit by Prompt Advisers / Mark Kashef, under the MIT license — a companion to Mark Kashef's Claude + Codex video. The guide is available in PT, EN and ES; the kit documents (README, GUIDE, PROMPTS) are in English. It is an independent educational resource, not an OpenAI or Anthropic product.
Open the Use Both guide →02 · How it all fits together
The courses and projects in this area are not separate pieces. Each one covers a stage, and all of them rest on the same thing: the project's Markdown files, which any runtime can read.
Claude → Codex course: why to separate the brain from the model. Codex Básico and Master Codex: master Codex. Codex Cheat Sheet: quick reference.
agente-claude-codex kit: audit, AGENTS.md, portable core (context/, tasks/, handoffs/), skills synced via polyskill, readback and handoff/prime.
Use Both: the 6 levels, the routing table, the prompts and handoff/prime with history.
claudex / iClaudeX: Claude writes PLAN.md, Codex critiques it from 3 angles, looping until LGTM or N rounds. MakeClaudeX: build plugins this way. DeepClaudeX: 70/20/10 multi-model.
The project's Markdown files — AGENTS.md, tasks, handoffs — which any runtime can read. The handoff/prime in agente-claude-codex and the one in Use Both are the same pattern; ours was hardened with the kit's rules: validated path, history that never overwrites and read-only prime.
Both get Claude and Codex working together, but at different layers: one is software that automates debating a plan; the other is a method for everything else.
| claudex | Use Both | |
|---|---|---|
| What it is | A Claude Code plugin (software) | A method kit: guide, prompts, templates and 2 skills |
| Coverage | Planning and review only (level 1) | 6 workflows: plan, image, build and review a diff, computer use, goal with a stop condition, handoff |
| How it runs | Automatic: Claude writes PLAN.md, Codex critiques it from 3 angles, looping until LGTM or N rounds | Manual or semi-automatic: you pass the artifact, or one calls the other through the CLI |
| Limit | N rounds, configurable | Recommends 2 at most; agreement is not proof, tests decide |
| Commands | claudex:plan, review, status, rollback, cancel, doctor | /handoff and /prime (Claude), $handoff and $prime (Codex) |
| State | Loop state, status and rollback | Dated snapshot in handoff/history/ + LATEST.md |
| Prerequisite | Claude Code + Codex CLI | None for the 5-minute trial; Python only for the installer |
A real case: someone with a Codex subscription who uses its image generator — image_gen, which at INEMA we call "image 2.5".
At INEMA today · script integration, no conversation
codex exec in non-interactive mode, asking image_gen for a PNG file in a folder.--gerador auto|codex|flux), the banner for project landing pages and guides (correct Portuguese text, which the local model gets wrong), scenes for the v6 courses and the area thumbnails of this very Events site.What the kit proposes · level 2, one piece at a time
The difference: ours is batched and automatic, with a local fallback; the kit's is manual and piece by piece, with more checking of cost and review.
03 · Every Codex + Claude project
Everything INEMA has published on the two tools: from the first command in Codex CLI to orchestrating several models, by way of migrating, staying agnostic and using both together. All open.
Migrate or stay agnostic.
Open the course → CourseThe complete Codex CLI course.
Open the course → (in Portuguese) CourseThe SaaS factory with AI agents.
Open the course → (in Portuguese) CoursePlanning with Claude + Codex in the terminal.
Open the course → (in Portuguese) CourseFrom prompt to production plugin.
Open the course → (in Portuguese) Course70/20/10 multi-model (GPT, Claude, DeepSeek).
Open the course → (in Portuguese)This area's sister: separate the brain from the model, migrate with the kit or stay agnostic for good. It is the migration path.
Open the area → KitA kit to migrate or stay agnostic: diagnosis, audit, portable core, skills and readback.
Open the guide → (in Portuguese) Kit · newThe six levels for using both together, copy-ready prompts and the handoff and prime skills.
Open the guide → CodeThe kit repository: GUIDE.md, PROMPTS.md, templates, the routing card and the skills installer.
Open on GitHub → CLIA CLI where Claude and Codex debate a plan.
Open on GitHub → ReferenceA Codex CLI reference.
Open the guide → (in Portuguese) Portal mapThe topic map on INEMA.CLUB.
Open the map → (in Portuguese)04 · Get started now
Pick a small task, write the brief and the acceptance criterion, ask one for the plan and the other for the critique, on the same file — two rounds at most. Then hand off before closing the session.
The six levels, copy-ready prompts and the handoff and prime skills, with a preview-first installer.
Open the guide →When the plan is big: Claude writes, Codex critiques from 3 angles, looping until LGTM or N rounds.
Open on GitHub →Before using both, separate the brain from the model: course, kit and method to migrate or stay agnostic.
Open the area →Not sure where to start? Talk to us on Telegram.