Audit before changing anything
Read-only inventory of both runtimes with doctor.sh and audit.sh: each skill comes out as reusable, adapter, native, or unresolved.
The problem
The lock-in isn't in the model. It's in what stayed inside the runtime.
Anyone who has used a coding assistant for a few months has accumulated instructions, memory, skills, hooks, and thousands of sessions in a format that only that runtime can read. The bill comes due on the day you switch.
- CLAUDE.md has grown, mixes rules with plugin leftovers, and no other agent knows what to do with it.
- Native memory and the JSONL sessions hold decisions that nobody wrote down anywhere else.
- Skills were copied by hand between runtimes, and the copies have already diverged.
- The Codex CLI has no native import: switching executors means rebuilding from scratch.
- A fresh session repeats old mistakes and contradicts what had already been resolved.
After the training
By the end of the 3 tracks, you can:
Separate the portable from the residue
Split CLAUDE.md: portable rules go into AGENTS.md, and whatever depends on a plugin or hook stays in the runtime-specific file.
Install the portable core
AGENTS.md, context/, tasks/current.md and handoffs/latest.md in the project, without overwriting anything that already exists.
Keep one source per skill
Canonical skills with polyskill: one source, per-runtime generated copies, and a drift command that shows what diverged.
Prove continuity
Readback in a fresh session on both runtimes, with the five continuity questions, and the daily handoff and prime cycle.
Add a third executor
Local model in a container, MCP and hooks between Claude and Codex, and a client workspace with explicit scope.
1 course · 3 tracks · 18 modules + the kit
Three tracks and the kit, from the why to your real project.
Track 1 defines each term the first time it appears. Track 2 runs the kit command by command. Track 3 applies everything in six projects on a real system. The kit closes the path: you run it on your own project, in audit mode.
- 1
Fundamentals
Track 1 · Why and what
Why separate the brain from the model, the vocabulary (runtime, harness, skill, MCP, hook, handoff), the three levels of migration, and the anatomy of a portable workspace.
- 2
Hands-on
Track 2 · The kit, command by command
doctor.sh and audit.sh, CLAUDE.md → AGENTS.md, the core installed without overwriting, skills with polyskill, and the readback proven on both runtimes.
- 3
Projects
Track 3 · Six projects on a real system
Migrate your first real project, MCP and hooks between Claude and Codex, memory curated with your approval, a local model in a container, and a client workspace.
- 4
Kit
agente-claude-codex kit
Scripts for diagnosis, adaptation, proof and drift, the portable core template, and mega-prompts A and B, all starting in MODE: audit.
Who it's for
Made for you if
- You have been using Claude Code for months and need to use Codex because of cost, access, or policy.
- You have already switched tools once and don't want to rebuild everything again.
- You want to run two executors on the same project, including a local model.
- You serve clients and need explicit scope and isolation per project.
Not for you yet if
- You want to use Claude and Codex together, one does the work and the other reviews: the Codex + Claude area covers that.
- You are looking for a button that copies everything from one runtime to the other without auditing.
Included in the training
The course starts from a kit that you run today.
The agente-claude-codex kit is bash and Markdown, open on GitHub. Nothing in ~/.claude or ~/.codex is copied in bulk or deleted, and installing a skill makes a backup alongside.
Projects and kits
Model: one context, several platformsagente-claude-codex kitagente-claude-codex · GitHubTopic mapCodex + Claude area
Videos
Why this training
What changes compared to copying CLAUDE.md.
Audit before implement
Every script and every mega-prompt starts by analyzing, planning, and simulating. You read the plan before changing a single file.
Proof by reading, not by file
A file existing proves nothing. The readback opens a fresh session in each runtime and checks whether the agent read and used it.
Ran on a real machine
The audit found 89 skills only in Claude: 71 reusable, 15 adapter, 2 native, and 1 without SKILL.md. The record is in the repository.
Three languages, same content
The English and Spanish versions are complete mirrors, with progress tracked separately per language.
Questions
Before you start.
Do I need to follow the order?
The suggested order is track 1 → 2 → 3 and then the kit on your project. If you already know runtime, harness, and handoff, you can start with track 2.
Migrate or stay agnostic?
Migrating means reaching level 1 or 2: importing into the app or running the kit. Staying agnostic is level 3, the only one that survives the next model switch. Track 1 helps you choose.
Will I stop using Claude?
No. Being agnostic doesn't mean abandoning Claude: you keep choosing the executor per task, with the knowledge living in the project.
Which tools do I need?
A terminal, git, and either Claude Code or Codex installed. The kit is bash and Markdown, with no interface. Module 3.4 adds a local model in a container.
How much content is there?
There are 3 tracks with 6 modules each, 18 in total, and 108 topics. Track 3 has six projects with an objective, commands, acceptance criteria, and risks.
Models change. Your structure stays.
Start with track 1 and pick a project you know well. In one session in audit mode you will already know what is portable, what needs an adapter, and what depends on a hook.