Claude → Codex area · model-agnostic workspace
Only CLAUDE.md and AGENTS.md are tied to the provider. Everything else you have built up — context, rules, decisions, tasks, skills, handoffs and memory — is portable Markdown. Once that material lives in a portable layer inside the project, Claude Code, Codex CLI, Gemini or a local model in a container become nothing more than executors: switching models becomes switching an edge, not the center. This area brings together the thesis, the method, the 3-track course, the script kit and the mega-prompts that make it happen in practice.
The course is open and available in English; the kit and the prompts are open and written in Portuguese. The live session has a date to be announced in the community.
01 · What is changing
The course starts from a simple discomfort: anyone who has used a coding assistant for a few months has piled up instructions, memory, skills, hooks and thousands of sessions in a format only that runtime reads. When a better, cheaper or mandatory executor shows up, the bill arrives. And it isn't a technical bill: it's knowledge stuck in the wrong place.
CLAUDE.md, native memory and JSONL sessions look like part of the work. They aren't: they are the storage of the work. The content is durable, the format is disposable — and mixing the two is what stalls a migration.
Rules, accepted decisions, tasks with a definition of done and handoffs are text. Written into files with an owner and a reading order, any agent that reads Markdown picks up where the other one stopped — including a local model running in a container.
You stop managing prompts and start managing context, state, tools, processes and validation. The model becomes a reversible choice: if it gets worse, gets pricier or disappears, the center stays in your repository.
Before · the brain inside the runtime
A giant CLAUDE.md · native memory · JSONL sessions · hand-copied skills · harness hooks
It works very well — as long as you only use that one runtime. Switching means rebuilding, and rebuilding means losing the decisions nobody wrote down.
After · the brain in the project
AGENTS.md · context/overview.md · context/decisions/ · tasks/current.md · handoffs/latest.md · skills with a single source
The names are a convention: no runtime loads these folders on its own. What makes them get read is the reading order at the top of AGENTS.md. That's why it works on any executor.
Models change. Your structure stays.
The thesis of the course, in five words.02 · The method
The kit's mega-prompts start in MODE: audit on purpose: first analyze, plan and simulate; only then implement. The rule that closes the method is the hardest one to accept — a file existing is not proof; the agent having read and used it is.
What exists today, without touching anything?
A read-only inventory of both runtimes: skills, commands, subagents, hooks and MCP. Each skill is classified as reusable, adapter, native or unresolved.
What is portable and what is residue?
CLAUDE.md splits in two: portable rules go to AGENTS.md; whatever depends on a plugin, a hook or a Claude menu stays in the specific file, which merely imports the portable part.
Can a fresh session carry on?
Five continuity questions — goal and definition of done, one rule with its source file, the last decision, the next action, conflicts — asked in each runtime.
What does the next session need to know?
Decisions, open items, next steps and paths in a Markdown file. The following session, in any runtime, reads it before acting. Handoff before closing, always.
Seven places, each with an owner and an update rule. This is what the kit installs without overwriting anything that already exists in the project.
superseded_by03 · Migrate or stay agnostic
The course splits the decision into levels of effort, not into opinions. Migrating means reaching level 1 or 2. Staying agnostic is level 3 — and it is the only one that survives the next model switch, which will happen again.
The shortest path when a ready-made import exists. Fast, and limited to what the other side accepts — the Codex CLI, for example, has no native import.
Run the agente-claude-codex scripts: audit, adapt instructions, install the core, port skills and prove it in a fresh session in both runtimes.
Reorganize the work so that knowledge lives in the project and the runtime is interchangeable. More work once; no migration afterwards.
Migrate now if…
Stay agnostic if…
There is no final state. Every provider change improves one executor and breaks another: closed models may do without skills that a local model requires, and a hook that exists in one harness doesn't exist in the other. That's why the method ends in maintenance — watching the drift between copies and logging every failure in one line, with the smallest possible fix.
Staying agnostic applies to accounts too. With two Codex subscriptions (personal and work, for example), the login stays outside the portable layer: each account lives in its own folder, pointed to by the CODEX_HOME variable (default ~/.codex). The project, the AGENTS.md, the skills and the handoff stay the same for both accounts.
# second account: its own folder, with the same instructions via links mkdir -p ~/.codex-trabalho ln -s ~/.codex/config.toml ~/.codex-trabalho/config.toml ln -s ~/.codex/AGENTS.md ~/.codex-trabalho/AGENTS.md CODEX_HOME=~/.codex-trabalho codex login --device-auth # prove it by reading, in a new session, on both accounts codex exec --sandbox read-only 'Read the AGENTS.md and list the 3 main rules' CODEX_HOME=~/.codex-trabalho codex exec --sandbox read-only 'Read the AGENTS.md and list the 3 main rules'
AGENTS.md, skills and handoff, shared by both accounts.auth.json, the history and the sessions. Never in the repository or in the handoff.CODEX_HOME changes. No project file changes..gitignore covers .codex*/ and that no script writes a login inside the project.CLAUDE_CONFIG_DIR variable separates accounts the same way.Caveats. codex login and logout without the right CODEX_HOME switch the main account. Do not copy the auth.json between folders: the login renews itself and two copies can invalidate each other. Services and terminals already open stay on the account they started with.
Want to use both together? Once the brain is separated from the model, you can use Claude and Codex on the same project — one does the work, the other reviews it. The Codex + Claude area brings together the six levels of the Use Both kit, who does what and every INEMA course and project on the two tools.
Open the Codex + Claude area →04 · The course
The course Claude → Codex: migrate or stay agnostic is open, with versions in English, Portuguese and Spanish. According to its own curriculum, it has 3 tracks, 18 modules and 108 topics; each module has six topics of about half an hour. Audience: people who already use Claude Code or Codex and want to use both — or not be locked into either.
It defines every term on first appearance: model, runtime, harness, skill, MCP, hook, handoff. This is the track that answers "why isn't copying CLAUDE.md enough".
Practice with agente-claude-codex: every module has at least one example you can copy and run, with the real commands from the scripts.
Each project has a goal, steps with commands, acceptance criteria and risks, using the diagnosis of a real machine as the case study.
Also in Portuguese and Spanish. The PT and ES versions are complete mirrors, with their own assets and progress tracked separately per language. Every page has the PT / EN / ES selector at the top.
Open the course →05 · The agente-claude-codex kit
The kit is bash plus Markdown files, and it can be used in three ways: the scripts, the mega-prompts (to run the migration by talking to an agent) and the template of the portable core, which you can copy into any project without running a single script.
doctor.sh answers "is my environment ready?" — each item comes out as ok, warning or missing, with the command to fix it. audit.sh writes a read-only report with the inventory of both runtimes and the classification of every skill.
The first one reads the project's CLAUDE.md and writes two proposals next to it, without overwriting. The second copies the portable core, listing what it created and what it kept. There is also faxina.sh, which classifies a long CLAUDE.md section by section.
One canonical source per skill, copies generated per runtime, and a drift command that tells you, per destination, whether it is identical or has diverged. Installing makes a backup alongside before overwriting.
It opens a fresh session in each runtime inside the project, asks the five continuity questions and saves the raw answer. Approval is yours, by reading the text: do the answers cite the right files, and does the next action match the task?
The two canonical skills that ship with the kit: one writes the handoff at the end of the session, the other reads AGENTS.md, context/, tasks/ and handoffs/ before acting and returns a briefing with the sources. Together, they close the cycle.
A weekly report on divergence between the skill copies and the global AGENTS.md; and a flow in which the agent proposes facts with source and date out of raw memory, and only what you approve enters the context.
What has actually been run on this machine (2026-09-13 and 2026-09-16). According to the kit's README: doctor.sh passed with two warnings; the audit found 89 skills existing only in Claude — 71 reusable, 15 adapter, 2 native and 1 without a SKILL.md; adapting the global CLAUDE.md separated 71 portable lines from 7 lines of residue; the core installed in an isolated clone passed the check; readback passed in both runtimes; the canonical skills ended up installed in four destinations with zero drift; and a fresh Codex session, after prime, pointed out three real contradictions between the handoff and the task. These are execution records in the repository itself, not a benchmark and not a measure of model performance.
Would you rather do it by talking? Mega-prompts A (migrate an existing Claude setup) and B (build a portable workspace, with personal and client add-ons) are available as copyable text in the repository. Keep MODE: audit on the first pass, read the plan the agent returns and only then run it in implementation mode.
06 · What INEMA already has on the topic
Fundamentals, hands-on work with the kit and six projects on a real system. With progress, notes and an export of the student's journey, all in the browser.
Take the course → Kit · step by stepLanding page and usage guide for the kit: diagnose, audit, adapt instructions, install the portable core, port skills and prove it in a fresh session.
Open the guide → (in Portuguese) CodeScripts, the portable core template, the canonical skills, PLANO.md with the analysis and the recommended pilot, and FALHAS.md with one line per failure fixed.
Open on GitHub → Mega-prompts A and BPrompt A takes selected parts of a Claude setup over to Codex. Prompt B builds the portable workspace. Both start in audit mode, plus the short readback and handoff prompts.
See the prompts → (in Portuguese) Topic mapThe topic page on the portal: understand, diagnose, migrate, prove and maintain — with the four executors on top of the same brain and the plan in phases.
Open the map → (in Portuguese) Other languagesComplete mirrors of the course in Portuguese and Spanish, with their own assets and progress tracked independently per language.
Abrir o curso →On the company side. If your question isn't which executor to use, but how to manage agents inside a company — roles, authority levels, cost per result and evaluation —, the AI Management area carries that conversation on.
Open the AI Management area → (in Portuguese)07 · Get started now
Pick a project you know well, run the diagnosis and read the audit before changing a single file. In one session you already know what is portable, what needs an adapter and what depends on a hook. The course and the kit are open; the live session about migrating or staying agnostic has a date to be announced in the community.
Three tracks: understand why to separate the brain from the model, run the kit command by command, and apply it to six projects with acceptance criteria.
Where the live session "migrate or stay agnostic" will be announced, with the real system on screen: audit, adapt, prove, handoff. And where you can discuss your own migration with people who have already done theirs.
Clone it, run doctor.sh, read the audit. Nothing in ~/.claude or ~/.codex is bulk-copied or deleted, and installing a skill makes a backup alongside.
Not sure where to start? Talk to us on Telegram.