He Reverse-Engineered 10 Apps in 10 Days and Got 6K Installs: He Published the Prompts He Gave Claude

Movez posted a workflow on X: use Opus 5.5 to reverse-engineer 10 projects in 10 days, with the apps collectively getting 6K+ installs and 33k+ visits. He also published a prompt he sent to Claude. First, draw the boundaries. REA's README states clearly: use only for lawful research, education, interoperability, software preservation, and analysis of software you own or are authorized to analyze; commercial tools such as IDA Pro, Hopper, and JEB require lawful authorization.

REA is an MCP service under morluto/rea, short for Reverse Engineer Anything. It lets an agent inspect local targets through tools, return findings and evidence, and then lets the agent explain behavior, write an implementation, and run tests. It covers native binaries, JavaScript/Electron, .NET, websites, Android APKs, firmware, SQLite, WASM, EVM bytecode, and more. Installation requires Node.js and npm; run npx rea-agents setup, select an agent, and restart. It supports Claude Code, Codex, Cursor, Gemini CLI, Grok Build, and others. Static JavaScript and .NET inspection does not require a native analysis engine; deep native analysis uses Hopper, Ghidra, IDA, or Kuna. Analysis runs locally, and results come with evidence and limitations. The README showcase includes DX-Ball sound panning calculations, Notion's Electron clipboard bridge, and TH04's DOS bullet-ring calculation. The pain point it solves: you see a feature and want to understand how it is implemented, you do not have the source code, and you need to trace the code, look at evidence, and then write your own version.

The prompt is far more than a single sentence: "help me write an app." The full version first requires reading the README, package.json, src/, and site/; every claim on sites, videos, and posts must come from code or the README, and features or numbers must never be fabricated. Then it splits into tracks: 0 writes a short plan, spawns parallel subagents, serializes tracks that touch the same file, and has each agent report changes, file paths, and verification methods; 1 does a UI/UX pass, explores the REA repository and inspora.design and 60fps.design, selects 3 to 5 patterns suitable for the current app, preserves the existing fonts and colors, and launches the app to screenshot dark and light before/after comparisons; 2 builds the launch website, referencing motionin.design, including a hero, an OS-detecting download button, three-step instructions, a privacy note, macOS and Windows installation notes, FAQ, GitHub star count and download count, footer, plus OG image, favicon, meta, 375px adaptation, prefers-reduced-motion, Lighthouse 90+, and deploys with npx vercel deploy --prod --yes; 3 creates a 20- to 30-second Apple-style motion video, in both 16:9 and 9:16 versions, with the pain point in the first 2 seconds, real UI screenshots and screen recordings, voiceover, word-by-word captions, and music; 4 writes 5 launch posts based on the video, 3 for X and 2 for LinkedIn, using only real numbers; 5 adds three tabs to the app: Dashboard, Tips, and Feedback. Tips connects to a Stripe widget, Feedback opens a prefilled GitHub issue, and Dashboard only does local statistics with no telemetry. The completion criteria are fixed: the site is live, the video is rendered, the 5 posts are written, the app builds and tests pass, and a report states what was done and what was not done.

What is most worth looking at in all this is the factual boundaries and acceptance criteria. Many AI demos stop once a page is generated. This prompt puts "only use facts from code and the README" at the very top, and puts screenshot verification, Lighthouse, build tests, and reporting what was not finished into the completion criteria. Reverse engineering sounds wild, but in practice it reads more like a project manager's checklist. 10 apps in 10 days sounds impressive; the truly time-consuming part may be waiting for Lighthouse and redoing screenshots.

The data needs to be viewed separately. The tweet says the apps collectively got 6K+ installs and 33k+ visits. Another screenshot lists dashboards for only four apps: cleanly had 7,990 visits, 3,039 downloads, 31% conversion, and 4,478 GitHub downloads; droneLab had 3,326 visits, 44,472 sessions, 1,142 quizzes, and 674 returns; hop had 4,010 visits, 1,006 downloads, 21% conversion, and 1,103 GitHub downloads; duoDesk had 3,144 visits, 664 clicks, 18 real-time online, and 913 GitHub downloads. The screenshot summarizes the four apps as 18,470 visitors, 4,709 downloads, 6,486 GitHub downloads, 4 apps, 4 dashboards, and 1,142 quiz answers. This is a sample of four apps; the total for the 10 projects does not appear in this image.

In the comments, someone said this pace would stress most development teams, and 6K installs shows the method works. Movez replied that he still had 20% of his weekly quota left. The attached image shows Ultra's Weekly · all models at 80% used, resetting Monday at 7:00 AM. The cost will land on subscription limits and verification time; do not just look at the headline of 10 apps in 10 days.

If you want to get started, do not try 10 projects first. Pick a target you are authorized to analyze, read the README and package, and have the agent use REA to trace one specific feature, requiring evidence and limitations; then write that feature as your own implementation and test it; only after that consider the site, video, and posts. The REA repository is at morluto/rea, the website is rea.tools, Movez's app list is at beyond-media.ai/projects, and the original tweet is here.