The author used the open-source project REA to reverse-engineer the Slides editor in Claude Slides, reproducing almost all of its functionality. The process: first download the core scripts and stylesheets from Claude Slides, then follow the instructions on the REA website to have Codex install REA, then use the REA Skill to reverse-engineer the obfuscated code, which yields code with a decent directory structure but a messy interior, and finally rewrite it with Claude Code 5.5 into well-structured, fully typed code.
The author reverse-engineers the Claude Slides editor with REA and rewrites the code, laying out a reusable five-step workflow and some prompt tricks.
I used REA to take a quick stab at reverse-engineering Claude Slides’ slide editor, capturing nearly all its functionality. (Figure 1)My approach was straightforward—not necessarily the best practice, but just for reference:
First, I downloaded the core scripts and stylesheet from Claude Slides. (Figure 2)
Following the instructions on the REA website, I had Codex install REA.
Then, I employed the REA Skill to deobfuscate the scrambled code. (Figure 3)
It didn’t take long before I ended up with a nicely organized directory structure—though the actual code inside remained quite messy.
Finally, I rewrote everything using Claude Code 5.5, resulting in clean, well-structured, type-safe code.
As for the prompt, instead of directly reverse-engineering it, I took a different tack: I first created a basic slide, then in the conversation, I told it I wanted to build a “slides skill” that could replicate what I’d made—and do so just as well. Naturally, this new prompt differed significantly from the original one, yet the outcome was virtually identical. (Figure 4 shows how I applied the same method to prompt Claude Dashboard.)For the complete code, see: https://github.com/JimLiu/baocut/commit/b8585e0c8d7a82bdee9f0ed019cf6edb075e41c1
REA: Let AI agents reverse engineer software for you when you don't have the source Open-source project REA (Reverse Engineer Anything) wires reverse engineering tools into AI coding agents like Claude Code, Codex, and Cursor. Spot a nice feature in someone else's app, and you can have the agent take the program apart, explain how the feature works with evidence, then write something similar in your own project. Reverse engineering here means working backward from a compiled program to figure out how it was written, without the source code. REA itself is an MCP server. MCP is the common interface AI agents use to call external tools; once connected, the agent can invoke REA's analysis features directly in the conversation. Installation is a single command, npx rea-agents setup, which registers it with Claude Code, Codex, Cursor, Gemini CLI, Grok Build, and other agents, backing up your config and asking for confirmation before changing anything. If you're not using an agent, you can also run the commands directly in a terminal. It can analyze a lot. Native programs (software compiled straight to machine code) rely on disassemblers like Hopper, Ghidra, or IDA to turn machine code back into assembly and C-like pseudocode. Electron apps (desktop software built with web tech, like Notion) can be unpacked directly from the ASAR archive in the installer, mapping out modules and inter-process communication without extra tools. It also supports .NET programs, Android APKs, firmware, websites, packet captures, Ethereum contract bytecode, and recording program runtime behavior on Linux and macOS. Analysis runs locally; REA doesn't upload the target program, but the results go to the model provider behind your agent, and how they handle the data depends on their policies. The project offers three case studies. First, the classic brick-breaker DX-Ball: starting from a single sound-effect call, it traced to the function that computes left and right channels from the ball's position, reconstructed incomplete pseudocode into C code, and 3205 test results matched the original, with the compiled 63 bytes identical to the original as well. Second, Notion desktop: tracing the full copy-paste path from the UI through the preload script and inter-process communication to the main process. Third, the game Touhou Fuumaroku (TH4) on the old Japanese PC-98: recovering the angle algorithm for the bullet ring from 16-bit instructions. For security researchers, people remaking old games, and software preservationists, where you used to read disassembly line by line and chase call layers by hand, you can now have an agent do the tracing first and verify the evidence it provides. For ordinary developers wanting to study how a competitor implemented some feature, there's another path. The README shows the project has 50000 stars on GitHub. The npm package first shipped in July this year, and the latest version, 6.3.0, was released today. Two things to note before use. The project states it only supports lawful reverse engineering research, and that authorization and compliance are the user's responsibility; reverse engineering someone else's software may violate terms of service or copyright law. It also states it has never issued or endorsed any cryptocurrency, and any token borrowing the REA name is unrelated to it. https://github.com/morluto/reaView quoted post on X
Source: 宝玉 · x.com