Source: Nate B Jones (AI News & Strategy Daily), raw/Your_Vibe-Coded_App_Can_Look_Right_And_Fall_Apart_After_Launch._REA_Shows_You_How_Real_Teams_Build..md (https://www.youtube.com/watch?v=PseF3LyXQ04) · Matthew Berman, raw/Everything_has_been_solved.md (https://www.youtube.com/watch?v=cX2kD2yQf88) and raw/Google_just_went_FRONTIER_Gemini_4_Argon.md (https://www.youtube.com/watch?v=M8iqahFogtA) · AI Explained, raw/AI_Labs_Warn_Race_Makes_Breach_Inevitable.md (https://www.youtube.com/watch?v=C4Sq3Rot8FU) · AI For Humans, raw/ai_put_minecraft_in_skyrim_and_open_ai_had_a_dev_day.md (https://www.youtube.com/watch?v=d09hcVXfXnU) and raw/Gemini_4_Argon_Meta_Muse_Wonder_Are_Good._Not_Great..md (https://www.youtube.com/watch?v=-OzVqLQVvgw) · All-In Podcast, raw/Is_Claude_Conscious_Pope_Rejects_Model_Welfare_Movement_OpenAI_s_Math_Backlash_France_Riots.md, raw/CoreWeave_Crash_AI_Trade_Cracks_Anthropic_IPO_Red_Flags_Bill_Gates_Doomer.md, raw/Dylan_Field_-_Is_AI_Replacing_Designers_Design_is_the_New_Code_Democratizing_Creativity.md · Robo Nuggets, raw/Someone_Just_Rebuilt_Adobe_for_Free_and_Claude_Can_Run_Them_All_7_apps_tested.md (https://www.youtube.com/watch?v=Fuo1i_-9Frc), for the ArtCraft repo link · GitHub, storytold org repo listing and READMEs (https://github.com/storytold/photocraft and sibling repos), read 2026-10-27 · AI Search, raw/AI_is_reverse_engineering_everything.md (“AI is reverse engineering everything”, https://www.youtube.com/watch?v=h5zkzon0gM4, uploaded 2026-10-06, about 14 min; sponsored by Runway), added 2026-10-30 for the tool list and repo links · MattVidPro AI, raw/Claude_Reverse_Engineered_Two_3DS_Games_to_Add_Real_3D.md (“Claude Reverse Engineered Two 3DS Games to Add Real 3D”, https://www.youtube.com/watch?v=o5iRKvXZEcY, fetched 2026-11-05), added 2026-11-09 for the 3DS plugins
Platform: YouTube (auto-captions, fetched 2026-10-21) · Creator commentary throughout; no first-party source for REA.
In October 2026 several channels reported the same new capability. AI Explained calls it “the newfound, likely unintended ability for you to use AI models to reverse engineer almost any piece of software out there.” It shows up in three forms: REA (“Reverse Engineer Anything”), a tool that lets Claude Code or Codex open software on your machine and explain how it works; game decompilation and mashups, where agents rebuild game code by guess-and-check against observed behaviour; and behaviour cloning, where an agent uses a commercial app and rebuilds it without seeing the source. The practical use for most readers is the first one: learning what your own software needs to do before you ask an agent to build it.
Key Takeaways
- REA is a requirements tool for non-developers. Nate B Jones calls it “an x-ray machine” for people who don’t build software and “an accelerator” for people who do. He reports “well over 100,000 GitHub stars in just a few days” (unverified).
- Behaviour you take for granted is the hidden spec. Windows Calculator’s % button gives 220 for “200 + 10%” and 20 for “200 × 10%”; Notion’s copy keeps text, HTML and a local structured-block entry linked by an identifier. “Make sure copy and paste works” is not a specification.
- Decompilation is a natural agent loop. Berman: “It has a loop, a verifiable loop. It guesses what something might be like, and then it can actually check it against the source material.” The output behaves identically but is not line-identical, because compilers discard names and comments.
- Speed-ups are dramatic but anecdotal. Chris Lewis’s Snowboard Kids decompilation took 84 days with AI, against “nearly 2 years” before. ^[ambiguous] The caption does not make clear whether the two years was his own earlier attempt.
- Most game mashups are illegal. AI For Humans: those using the original games’ textures and sound effects are not legal in the US or EU. A clean-room rebuild is “kind of a gray area.”
- Stay client-side. Nate: you will not see server-side code, “nor should you try because that would be hacking.”
- The tooling is public and agent-ready (added 2026-10-30). AI Search links Universal Modder, REA, Ghidra and IDA MCP servers, ILSpy and Cpp2IL; you paste the repo URLs into the agent’s prompt. See the modding-toolkit section.
- Sharbel’s take (added 2026-11-01,
raw/These_10_Open_Source_Tools_Should_Be_Illegal.md). He ranks an autonomous reverse-engineering plugin for Claude Code or Cursor, most likely REA ^[inferred] since the captions don’t name it, #1 on his “should be illegal” list. He says it recently hit #1 on GitHub trending, is “in beta”, is “one developer”, and “inherits every mistake the underlying model makes”. Pointing it at systems you don’t own “is how you get into real trouble fast.” See Ten Public Security and AI-Fakery Tools. - A non-technical playbook built on REA’s loop (added 2026-11-13,
raw/Reverse_Engineer_Anything_Does_Way_More_Than_You_Think.md). Austin Marchese reduces REA to five steps (pick, gather, design, rebuild and grade, iterate) and applies them to competitor funnels, single features, your own repeated tasks and expert thinking. He rebuilt Apple Notes in about 40 minutes but calls whole-app clones “genuinely useless”. See The REA Five-Step Playbook. - “If your only moat is the actual software, I think you’re going to be in trouble” (Berman). The All-In hosts split on how far that goes (see the moats section).
REA: learning how real software is built
What it does. REA plugs into Claude Code or Codex and “lets your AI open up software on your computer and show you how it’s actually built” (Nate B Jones). It needs Node.js and npm, and different targets need different inspection tools.
Worked cases (from REA’s published investigations, as described by Nate):
| Target | What the investigation found |
|---|---|
| Aegis (Android authenticator) | The code combines an account key with the current time grouped into 30-second intervals. A demo lets you move the clock and watch the code change. |
| Notion copy and paste | Text and HTML go to the clipboard; structured block data stays local, with an identifier in the HTML so paste can reconnect them. |
| Windows Calculator % | After + it becomes a percentage of the first number; after × it uses the operator already chosen. |
| Chromium dino game | Speed starts at 6 and adds 0.001 per collision-free update while below 13. |
| DX-Ball | A brick’s horizontal position becomes a left-right stereo pan value, recovered despite incomplete decompiler output. |
Ledger Reborn. A community developer investigated QuickBooks Desktop for Mac with REA (inventory arithmetic, budget signs, reconciliation balances, cash-basis branches), then independently implemented an accounting app in Python and SQLite. Nate’s point is that copying a screen reproduces where a number appears, not how it was produced; the investigation went into the calculations, “where the software’s meaning actually lives.”
Why it matters for vibe coders. Nate’s examples of under-specified features: a reminder agent that was late because “the clock started late,” and “approve a request” (which permissions change?) or “undo” (how much work should disappear?).
Game decompilation and mashups
- The loop (AI For Humans’ Kevin Pereira): press a key in the original, run the candidate engine, compare jump, tilt and velocity, and iterate “across every system like inventories, sound effects, movement, physics” until the result is “frame perfect.”
- Four mashup techniques (Berman):
- Asset swap: one character’s skin on another’s behaviour.
- Pass-through mod: both games run and a bridge syncs cameras, drawing and events (Minecraft TNT exploding in GTA).
- Mechanic transplant: carry acceleration, jump height and collision rules into another engine.
- Full rebuild: write a replacement for the game’s engine and rules.
- Examples reported: a Skate 3 × Modern Warfare 2 mashup built with Claude Code and Opus 5.5; a fal.ai creative technologist’s “universal modder” on GitHub, built while back-building Terraria and Age of Empires 2 (AI For Humans). Berman has GPT-6.1 Sol rebuilding Super Mario World for the ModRetro handheld from documentation and screenshots; it had run “well over 4 days.” He adds, “I have resets backed up that I haven’t used.”
- Parallelism is the speed-up. Berman: give chunks to many agents at once.
- Stereoscopic 3D for two 3DS games (MattVidPro, added 2026-11-09). Claude Code reverse-engineered Minecraft and Terraria on the Nintendo 3DS, “reading the machine code to work out how these games draw pixels, and then patching them in real time as they run”, to add true stereoscopic 3D. He says he “didn’t write a line of the code” and supplied “the vision, the testing, the direction” (
raw/Claude_Reverse_Engineered_Two_3DS_Games_to_Add_Real_3D.md).- The loop. Claude first searched open-source projects for prior attempts and found one that had dug into Minecraft’s code, then iterated in an emulator. He play-tested on real hardware while recording himself narrating, and sent the resulting MP4 back: “That video file literally became the prompt.” Plugins that loaded in the emulator often crashed on hardware, because emulators do not simulate the hardware at full precision.
- The rendering trick. Rendering each eye separately ran at “around 17 hertz” per eye, doubled CPU drawing and doubled game logic (rain fell “twice as hard”). The fix: the game builds each frame once, and the plugin re-submits the left-eye frame to the GPU for the right eye with “a small sideways shift”, so CPU work is almost unchanged and the GPU, which had spare capacity, draws twice.
- Self-audit. Each pass ended with AI auditors reading the code for crashes; “the auditors caught real bugs all the time.”
- Effort and release. “Somewhere between 12 and 16 hours of pure development time with Claude”, spread over a few days. Both plugins are on GitHub under the MIT licence and contain no original game files; players must supply their own legally acquired copies.
The modding toolkit, with links (AI Search, added 2026-10-30)
AI Search’s walkthrough is the first source in this article to name the repositories. Its framing: GPT-6 Astra’s launch page shows over 99% success after four attempts on SRE-Bench, a benchmark of reverse-engineering software from binaries, and “Claude Opus 5.5 appears to be even better” (his impression, no figure). This sits alongside the 88.0% in one attempt recorded in the GPT-6 article: the two are different attempt counts, not a conflict.
Three approaches, in rising difficulty:
- Pass-through mod. Both games run; the agent writes a bridge that translates events (Minecraft TNT placed → GTA spawns an explosive at that position). Easiest, but “kind of like duct tape”: you stay inside both original engines, so collision and physics glitch.
- Rebuild the game. The agent runs the binary through a decompiler such as Ghidra, works out animations, maps, models and mechanics, and rewrites the game, often in Rust, which he says frontier models are good at. You need to own a copy. This also enables ports: VR mods of decompiled Super Mario Galaxy and Mario Kart for Meta Quest, and PS5 games running on PC.
- Mechanic transplant. Copy one system into another game: Mirror’s Edge parkour into Skyrim (“done in around a day”), Skate 3 trick physics into Call of Duty or World of Warcraft.
Repositories he links (give the URL to your agent in the prompt):
| Repo | What it is |
|---|---|
github.com/rehan-remade/universal-modder | Universal Modder, with a mashup skill for pass-through mods and porting content between games |
github.com/morluto/rea | Reverse Engineer Anything (REA), pre-built skills and MCPs for the agent |
github.com/trevaintdead/ai-game-modding-guides | Step-by-step guides: a full pass-through mod and a Rust rewrite, plus a prompting and workflow section |
github.com/bethington/ghidra-mcp | MCP server for the Ghidra decompiler |
github.com/HexRaysSA/ida-mcp | MCP server for IDA |
github.com/icsharpcode/ilspy | ILSpy, for .NET games and apps |
github.com/SamboyCoding/Cpp2IL | Cpp2IL, for Unity games |
github.com/boykopovar/AnyPS5, github.com/yuriolive/PortPS5 | Projects using AI to make PS5 games run on PC |
Prompting: “there are no magic prompts.” With Opus 5.5 in particular, a plain request (“I want to put this character from this game into another game”) plus the MCP links is enough, per the presenter. Models he names for this: GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5.
Limits he states: he will not show cracking commercial software “because there are, of course, legal implications,” and closes with a sarcastic aside about paying for Adobe subscriptions. His speculation that labs are reverse-engineering all the world’s software for training data is unsupported.
Behaviour cloning of commercial apps
- PhotoCraft: an open-source Photoshop clone, made by pointing an agent at a local Photoshop install and having it use and recreate the app (Berman).
- Excel: Berman’s earlier agent-built Excel clone worked by clicking and observing. Gemini 4 Argon, given the task without a loop, opened Excel on his computer and produced a working browser version in about 20 minutes, with a formula-entry bug.
- Adobe suite: Chamath Palihapitiya (All-In) says people “decompiled, distilled… and published open-source versions” of Adobe’s five main products in Rust. AI Explained mentions “observing software behavior to recreate the entire Adobe suite.”
- Slow Books Pro: a free, local, forkable QuickBooks clone for Windows and Mac (AI For Humans).
What this does to software moats
Cross-source synthesis; the positions are the hosts’ opinions.
- The bear case. Chamath: “IP in software was effectively rendered worthless,” because once services go headless and are wrapped in agent tools over MCP, “the value of IP does no longer exist in software.” Berman: add services (training, enterprise features, support, maintenance) if the software is your only moat.
- The counter-case (Jevons for software). David Sacks: coding assistants create “this vast new category” of bespoke software that was never worth building for a mass market, and “demand for coders is increasing even though 95% of the code’s written by coding assistants.”
- Moats the All-In hosts still name: proprietary data with vertical feedback loops (Harvey’s law-firm data is the example given), and power and land for compute.
- A data point against “SaaS-pocalypse.” On All-In, a host notes Figma grew 48% year over year in the quarter after Anthropic’s Claude Design announcement; Dylan Field did not dispute it.
- The legal limit. AI For Humans: US and EU law protects expressions that match an original, which is why asset-reusing mods are illegal; the hosts expect many to be shut down.
- Implication for builders ^[inferred]: defensibility moves to what an agent cannot observe from the client: server-side logic, accumulated customer data, distribution, and service.
Try It
- Before your next vibe-coded feature, investigate one app that already does it. Pick the one behaviour you care about (copy-paste, a reminder, a percentage) and ask the agent to trace only that.
- Tell it how deep to go. Nate: “Tell your microscope what to focus on.” A whole-app explanation buries the one decision you need.
- Ask your agent whether each optional REA install matters for your target before installing it.
- Turn the findings into acceptance tests ^[inferred]: 200 + 10% = 220 is a test case, not trivia.
- For a rebuild, give the agent a verifiable goal and a loop (Claude Code, Codex and Cursor all have
/goalor/loop, per Berman), plus reference footage or screenshots to compare against. - For a decompile-and-rebuild on a game you own, connect the agent to the Ghidra or IDA MCP server (or ILSpy/Cpp2IL for .NET/Unity) and start from the ai-game-modding-guides repo’s worked examples.
- Do not reuse another product’s assets, and do not probe server-side code.
Two Minute Papers’ summary of the research (2026-11-11)
Károly Zsolnai-Fehér cites, without naming the papers:
- a reconstruction of one program that “only passed 15% of the test” and “in just 9 months, it jumped to nearly 100%”;
- a 12-million-parameter model recovering equations from software without source;
- web-page reconstruction at 0.938 visual similarity.
His reasons it works: open-source repositories supply unlimited code-and-app training pairs, and the original program acts as a test oracle. The papers are unnamed, so none of the figures were checked.
Open Questions
- REA’s repository, licence and real star count. No source gives a URL; the 100,000-star figure is Nate’s.
- Partly answered 2026-10-30: AI Search links it as
github.com/morluto/rea. Licence and star count are still unchecked. - Install command (2026-11-13): a search-result snippet of the README shows
npx rea-agents@latest setup(page not fetched).
- Partly answered 2026-10-30: AI Search links it as
- Legality of decompiling commercial binaries beyond asset reuse: no source addresses licences, end-user agreements or the DMCA in detail.
- The Adobe Rust clones and PhotoCraft are reported, not linked. Feature completeness is unknown.
- Answered (2026-10-27): the clones are ArtCraft’s, in the GitHub org
storytold(display name “ArtCraft”; Robo Nuggets gives the linkgithub.com/storytold). There are seven Adobe counterparts, not the five Chamath mentioned: PhotoCraft (Photoshop), FilmCraft (Premiere), EffectCraft (After Effects), VectorCraft (Illustrator), DesignCraft (InDesign), LightCraft (Lightroom) and PdfCraft (Acrobat). Each repo is Rust, licensed Apache-2.0 (PhotoCraft’s README says MIT or Apache-2.0), and was created 2026-09-30 or 2026-10-01. PhotoCraft’s README calls the code clean-room, “implemented from public specs and observed behaviour only.” On completeness it says PhotoCraft is “early alpha” and “not yet a Photoshop replacement for daily professional work.” It lists the biggest gaps as AI/generative features, 15 missing tools, depth in typography and pro workflows, and plug-in compatibility. Every Photoshop menu item is wired to a command, but the README says that “measures wiring, not behaviour.” DesignCraft’s README gives its stage as pre-alpha, and the other READMEs carry badges such as “young and moving fast.” Two things are still unverified: Berman’s account of an agent recreating PhotoCraft from a local Photoshop install, and Jay’s claim that the apps were written with Opus 5.5. The README doesn’t describe how the code was written. Fuller coverage: Claude Code Driving Creative Tools.
- Answered (2026-10-27): the clones are ArtCraft’s, in the GitHub org
- The Snowboard Kids timeline (84 days against “nearly 2 years”) is ambiguous in the caption.
Related
- Ten Public Security and AI-Fakery Tools (Sharbel) — REA ranked as the #1 tool, alongside voice cloning and AI-label stripping
- Software Factory Bottleneck Playbook — once code is cheap, specification is the bottleneck
- Loop Engineering — the verify-and-iterate loop decompilation runs on
- Verifier-First Loops — reference footage and screenshots as the judge
- The Gauntlet Loop — blind reviewers comparing a rebuild against the original
- The Verification Frontier — why checkable tasks fall first
- Product After Cheap Code — product judgment once building is cheap
- Claude Design — the launch behind the Figma growth data point
- Ben AI’s Skill-Engineering Framework — the same reverse-engineering idea applied to skills.