Learning path map
Detailed content
🧭 Reverse engineering is not a product
What the videos about mods showed, why the workaround is only useful for discovery, and which official routes do the same thing today.
Videos showed Claude Code using the Codex app's computer-use feature and several sessions in parallel, almost entirely through reverse engineering.
Knowing what was done and how it was done separates the idea (good) from the approach (fragile).
Claude Code, Codex, mod, computer use, reverse engineering.
The course rule: reverse engineering proves something is possible; the day-to-day system uses the official route.
Anything that depends on a makeshift connection breaks with the next update, and only you know how to fix it.
Discovery, product, contract, fragile label.
The switching table: codex exec, claude --bg, and claude agents, official plugins, and written permissions instead of tricks.
Each table row becomes a course module that you run on your machine.
Official method, CLI, background session, plugin.
The INEMA Agent Runtime: a runtime/ folder with four rules files, scripts, bridges, a guard, and seven tested recipes.
Works through Claude Code and Codex subscriptions, with no paid API or server to maintain.
Copyable kit, subscription, quota, API, recipe.
Discover, connect, route, execute, observe, verify, and learn, with one kit file for each step.
The cycle follows the course order and the order in which you connect agents to your work.
Cycle, policy, proof, command → expected output.
The clinic owner with her schedule in a spreadsheet and the accountant with an ERP that has no API, whose data comes in the kit itself.
With the same example files, what you run matches what the lesson shows.
ERP, CSV, example files, final project.
🩺 Install and diagnose
From zero to "READY": prerequisites, cloning the kit, the doctor.mjs diagnosis, and the first request to the agent.
Node 18 or later, Claude Code and/or Codex signed in with a subscription, Linux or Mac (Windows via WSL).
Without these three items, no recipe runs; with them, they all run.
Terminal, Node, WSL, subscription sign-in.
Download the kit repository to a folder of your own with one command.
The cloned folder becomes your project: rules, bridges, and recipes are already in place.
Git, repository, clone, project folder.
A script that checks Node, Claude Code, Codex, and Ollama and ends with READY or MISSING.
It only reads versions and status: it doesn’t call a model or use up quota.
Diagnosis, exit 0, READY, MISSING.
What each status means and the command the doctor itself suggests to resolve it.
You fix what's missing yourself, without guessing.
ok, missing, not logged in, outdated, optional.
Codex reads AGENTS.md; the kit's CLAUDE.md points to it. The same rules apply to both.
You write the rule once and don't have to maintain two versions.
AGENTS.md, CLAUDE.md, @AGENTS.md, .claude/settings.json.
Open claude in the kit folder and ask it to read LEIA-ME and help fill in CAPACIDADES.md.
It’s the first real use: the agent reading the kit’s rules before acting.
Interactive session, trusted folder, first prompt.
🪜 The ladder of pathways
Seven levels, from API to reverse engineering: how to choose the most stable route and require a test before accepting a “can’t do it.”
The LEIA-ME table: API, MCP, CLI, SDK, computer use, local bridge, and reverse engineering, with examples and stability.
It’s the most useful decision rule in the kit, and it fits on one page.
Via, level, API, MCP, CLI, SDK, local bridge.
Each rung has a stability level: how well it keeps working when the other side changes.
The route that breaks less costs less to maintain, even if it takes more work on the first day.
Stability, maintenance, contract, fragile.
The LEIA-ME and AGENTS.md rule: before concluding that something is impossible, the agent tests each level with a command.
Often the route exists; it just wasn’t documented.
Premature impossibility, concrete test, evidence.
The ERP without an API goes step by step all the way to CSV export, the local bridge.
It’s the most common case in a small office: an old system that only exports a file.
Export, CSV, local bridge, read-only.
The spreadsheet schedule also uses the local bridge; reading is straightforward, writing requires approval.
Shows that the same route can have different policies for reading and changing.
Spreadsheet, agenda.csv, read × change, N2.
A prompt for the agent to walk you through the ladder, testing each level on your system.
You finish knowing which method to use in the final project.
Path diagnosis, evidence by level, choice.
🗺️ The capabilities map
CAPACIDADES.md: one line per system, with access method, level, how to call it, policy, and test date.
System, route, level, how the agent calls it, policy, and tested on.
Each column answers a question you’d ask before letting someone use your system.
Capabilities map, route, policy, test date.
Codex CLI through the codex-exec.sh bridge and Claude Code in planning mode, both with read-only policies.
These are examples of what a well-written line looks like.
Level 3, read (N4), tested line.
The file’s four example rows, all marked as pending until tested.
Starting with an example similar to your case cuts the work in half.
Pending, local bridge, computer use, N2.
The README prompt for the agent to fill in the map with you, one system at a time.
The agent knows the ladder; you know your work. Together, you can map it out in minutes.
Guided interview, one question at a time, with human review.
The file rule: a system without a line is not used, and a line without a test stays pending.
Prevents the agent from making up access to a system that no one has checked.
Allowlist, test, date, pending.
Check what the program already offers (export, integrations, command line) before thinking about shortcuts.
The good side of reverse engineering is that you discover the path without relying on a hack.
Observe, export, integration, LIMITES.md.