5 Days, 1.7k Stars: This REA Prompt Library Turns 'How Do I Ask?' Into Copy-Paste

REA (Reverse Engineer Anything) is an open-source MCP server and CLI. Its author morluto uses it to give AI coding agents a real disassembler, letting the agent read all the way down to the binary layer. REA hit No. 1 on GitHub Trending on October 8, and Perplexity then made it a built-in skill for Computer agent. Replit CEO Amjad Masad called this wave of AI reverse engineering 'absolutely insane; soon all software will be de facto open source.' But ten days later, most people who starred REA are still talking to it in a single chat window. A single session can handle one function fine, but by the 40th function it starts to drift: naming structs repeatedly, treating old guesses as facts, and finally using the decompiled code it just read to verify its own reconstruction.

That is exactly the problem REA-Prompts (REAP) aims to solve. This prompt library, introduced in a tweet by Gyomei, got 1.7k stars in 5 days, and its core is turning 'how should I ask REA' into copy-paste. Replace the content in angle brackets and keep each request focused on one function. The project has 17 sections covering function tracing, binary analysis, software cloning, game logic recovery, WASM porting, firmware, Android APK, kernel, Electron, vulnerability research, anti-cheat, and DMA defense research, packaged into 26 skill packs. Every prompt enforces the same set of rules: conclusions must include the file, function, and address; confirmed facts and guesses must be marked separately; ask about only one function at a time. The project README repeatedly emphasizes: the narrower the question, the more trustworthy the answer.

REAP lowers the barrier to asking questions, but what is really interesting is Gyomei's 7-step guide. He pulls REA out of the chat box and splits it into 4 sub-agents, each running in an independent context:

• Analyst: opens the target through REA, maps it out, and marks every conclusion as observation, inference, or unknown.
• Spec Writer: looks only at the analysis results and outputs a pure behavioral spec, with no addresses, offsets, or decompiled names.
• Rebuilder: in an independent container, writes code only from the spec, and REA is not in its tool list.
• Verifier: runs the original program and the rebuilt version separately on the same set of inputs, and reports the match rate.

The key to this design is the third line: rules in the prompt can be ignored, but if REA is not in the tool list, it really cannot be called. The Rebuilder has no access to the binary, and reverse engineering and reconstruction are strictly isolated. REA officially demonstrated verification of the DX-Ball sound function: all 3205 original x86 test cases passed, and all 63 compiled bytes were reproduced. Gyomei cited this case in the guide as a data point for the method.

There was plenty of skepticism in the discussion. One commenter said, '1.7k stars, and maybe only four people will actually run a prompt,' to which Gyomei replied, 'Then only four people will surpass everyone else.' Another commenter said, 'A prompt library is only useful if it includes failure cases,' and 'Match rate is everything; a number without difference details is not credible.' These criticisms point to the same issue: the prompt library and workflow are only the first step. Producing verifiable results and publishing the differences is what truly earns trust.

If you want to get started, the starting point is simple: choose a program you wrote yourself, and pick one function you can explain in a single sentence. Electron and .NET programs are the easiest, and REA does not need a native analysis engine. If you want to use the full workflow, install REA according to Gyomei's guide, put the four agents—Analyst, Spec Writer, Rebuilder, and Verifier—into their respective folders, and then run a script. Note: REA's official disclaimer limits use to legitimate research, interoperability, software preservation, and software you own or have the right to inspect; licenses may also prohibit reverse engineering. After one round, record the match rate, number of rounds, and token consumption. These numbers are more persuasive than any opinion.