auto-reverse-engineer

auto-reverse-engineer is a prompt framework for running autonomous reverse-engineering projects in an existing coding-agent harness.

Files

• bootstrap.md: creates a new project workspace, asks for the required context, and prepares the run
• program.md: runs the actual reverse-engineering loop inside that project workspace

How it works

1. Run an agent with bootstrap.md.
2. It creates a separate project folder for the target.
3. It writes the initial files such as goal.md, project-context.md, progress.md, attempts.md, paths.md, inbox/, and knowledge-base/.
4. It copies program.md into that project folder.
5. Start a second agent in the new project folder with program.md.

Running it

From the repo root, first run bootstrap.md to create the project workspace. In this example we use Codex, but the same can be done with Claude or any other agent harness.

codex "run bootstrap.md"

It will ask you for needed information and create the project workspace.

After bootstrap, start the main run from inside the created project folder: For Codex, the agent does not always keep looping by itself, so run it in a shell loop from inside the project folder:

while true; do
    codex exec --yolo "run program.md, target projects/<project-slug>/goal.md" 2>&1 | tee -a agent.log
    sleep 1
done

Claude usually does not need a shell loop, so you can just run it once:

claude -p "run program.md"

Project model

Each target gets its own workspace. The framework repo stays generic. The project folder holds the target-specific state, artifacts, notes, scripts, and logs.

The runtime agent should:

• work toward the goal in goal.md
• keep a structured knowledge base
• track attempts and avoid repeating failed work
• maintain paths.md with prioritized investigation paths, blockers, and trigger conditions
• poll inbox/ each loop to detect when the human provides requested resources
• automatically unblock and pursue paths when trigger conditions are met
• always work on the highest-priority ready path

Example

A project could target a CGM app and sensor:

• inspect the APK
• recover the BLE protocol
• identify pairing and message parsing
• reach raw glucose data from the device

Purpose

The point of this repo is simple: make autonomous reverse-engineering runs repeatable, structured, and isolated per target.