Claude Code rebuilt a paid app's main feature in 47 minutes. Here are the prompts to see how an app on your PC is built.
In the run behind the reel, Claude Code read an AirPlay screen mirroring app installed on Jason's PC, concluded from the names and strings in its files that its engine was a port of the open-source UxPlay, and built a new Windows program around UxPlay. This guide gets you the first half for an app you have installed: a written report on what the app is built from, with the evidence for each finding. If the report shows a way, one more prompt has Claude Code build your own version of the main feature, from new code and open-source parts. The tool is REA (Reverse Engineer Anything), an MCP server and skill for Claude Code.
The guide covers what REA is, what to have before you start, the four prompts for Windows and Mac, what Jason's run showed, and the fixes for a run that stalls.
02Paste the report prompt for an app you have installed
03Read the report, then decide whether to build your own version
To paste into Claude Code, the last one optional
4 prompts
The last two are optional
6 steps
From the prompt to a built program, in Jason's run
47 min
The license on REA, the MCP server and skill
MIT
What you will have when you finish
A report on how one app you have installed is built, and the option to build your own version of its main feature.
REA connected to Claude Code: an MCP server that gives Claude Code tools for inspecting software, and a skill that tells Claude Code how to use them
REPORT.md for one app you have installed: what it is built from, which open-source projects it ships, and how its main feature works, with the evidence for each finding
A plain statement of what REA could and could not see on your computer, and a list of what could not be determined
A way to tell whether a build is worth trying, taken from what made the app in the reel a simple case
Optional: your own version of the main feature, built from new code and open-source parts, with self-tests and an end report that names what is untested
Six steps, the last two optional
You use Claude Code and know the basics. You have never added an MCP server or a skill. These steps are written for that reader. REA is both: an MCP server gives Claude Code tools it can call, and a skill is a set of instructions for using them. Jason's run, the one in the reel, started from one short prompt, and Claude Code made eight judgment calls to fill the gaps. The prompts below state those calls up front. They were written from that run and have not been run again as written.
Claude Code, signed in, and Node.js with npm: version 22.19 or later, 24.11 or later, or 26 or later. Check with node -v. You also need the app installed, and it must be an app you own or are authorized to examine. The optional build needs more: a compiler (for a C++ app, Visual Studio Build Tools 2019 or later with the C++ workload), git, an internet connection and free disk space. Jason's build used about 3 GB. Tested on Windows with Claude Code v2.1.296 and REA 6.4.0.
01
Install REA and connect it to Claude Code. This prompt is REA's own: it has Claude Code run REA's setup, show you the setup plan before anything changes, and check the result. Setup adds one entry named rea to Claude Code's configuration file, C:\Users\<you>\.claude.json, and makes a backup of that file first. REA's docs say it also puts the REA skill in C:\Users\<you>\.claude\skills.
Install REA and connect it to this coding agent using npx rea-agents@latest setup. Show me the setup plan for approval, then verify the installation.
Your part: approve the commands Claude Code asks to run, read the setup plan, and approve it. If setup asks which agents to connect, choose Claude Code. Then exit Claude Code, make a new, empty folder for this work, and start Claude Code again inside that folder. REA's tools do not appear in a session that was already open. Type /mcp and check that rea is in the list. No rea? Run node -v. REA runs on Node.js 22.19 or later, 24.11 or later, or 26 or later, and does not run on 23 or 25.
02
Find the app's install folder. This prompt asks Claude Code to look up one installed app, report its name, its version and its folder, and change nothing. For a Microsoft Store app, the folder is C:\Program Files\WindowsApps\ followed by the package's full name.
Paste into Claude Code, with one word from the app's name swapped in
Find where one app is installed on this PC. Change nothing.
One word from the app's name: <one word, for example AirPlay>
1. Run this command in PowerShell. Put my word where AirPlay is. Show me the output:
Get-AppxPackage *AirPlay* | Select-Object Name, Version, InstallLocation
2. That command matches package names, and a package name can differ from the name on the Start menu. If the output is empty or shows a different app, run this and show me the output:
Get-StartApps | Where-Object Name -like "*AirPlay*"
The part of AppID before the exclamation mark is the package family name. Then run:
Get-AppxPackage | Where-Object PackageFamilyName -eq "<package family name>" | Select-Object Name, Version, InstallLocation
3. If both lookups are empty, say so. Then look for an app that did not come from the Microsoft Store, and tell me how you found it.
4. Tell me three things: the app's name, its version, and the full path of its install folder.
5. List the files at the top level of that folder, with their sizes.
Do not copy, move, change or delete any file. Do not install anything.
Your part: swap in one word from your app's name, then copy the name, the version and the folder path from the answer. You paste all three into the next prompt. On Jason's PC both lookup commands ran without administrator rights. His run covered one Microsoft Store app, and other kinds of app are untested.
03
Paste the report prompt. It tells Claude Code to study the app with REA first and to say which decompilers REA can use on your PC. A decompiler turns compiled code back into code a person can read. If REA has none, Claude Code falls back to three things it can read from the app's files without one: imports (functions a file takes from other files), exports (functions it offers) and strings (readable text inside it). It writes the findings to REPORT.md with the evidence for each one, builds nothing, and leaves the installed app alone.
Paste into Claude Code, with the three blanks filled in
Use REA to study one app that is installed on this computer. Write a report on how the app is built. This is a read-only study: build no program. You may write small helper scripts in the current folder to read the app's files.
THE APP
- Name: <app name>
- Version: <version>
- Install folder: <full path of the install folder or app bundle>
I own this copy and I am allowed to examine it.
WHERE TO WORK
- Work only in the current folder. Put every file you write here.
- Never write into the app's install folder. Never change, move or delete the installed app.
- Read the app's files where they are. If a tool needs its own copy of a file, put the copy in a subfolder named evidence and tell me. That subfolder stays private.
- Install nothing. If a tool you want is missing, say so and use what is already on this computer.
USE REA FIRST
1. Check that you have REA's tools. Their names start with mcp__rea__. If you have none, stop and tell me that REA is not connected to this session.
2. Call REA's binary_session tool. Tell me which decompilers REA can use on this computer. REA calls them providers: Hopper, Ghidra and IDA. Say plainly whether REA can decompile native code here.
3. Open individual files with REA. Its open_binary tool accepts a file or a macOS app bundle and rejects any other folder.
4. Use the REA tools that fit these file types. You do not need to try every tool.
5. Record each open question with REA, then export REA's evidence bundle into the current folder.
THEN THE FALLBACK
If REA cannot decompile native code here, keep going. For each main file, list its imports, its exports and its readable strings with ordinary tools. Label every finding with its source: "REA" or "ordinary tools". Only REA output counts as REA evidence.
WRITE REPORT.md WITH THESE SECTIONS
1. What REA could do on this computer, and what it could not see.
2. The files: each main file, its size and its job.
3. What the app is built from: languages, frameworks and libraries.
4. Open-source projects the app ships or is built on. For each one: its name, its license, and the strings or names that point to it.
5. How the main feature works, step by step, as far as the evidence goes.
6. License checks, accounts and servers: say whether the app has each one. Report what you see and bypass nothing.
7. What a rebuild of the main feature would need: the parts that exist as open source, the parts that are missing, and the build tools.
8. What could not be determined. For each item, say what would settle it.
RULES FOR EVERY FINDING
- Mark it "observed" or "inferred".
- Give the evidence: the file name, plus the string, import, export or REA record behind it.
End your reply with a five-line summary and the full path of REPORT.md.
Your part: fill in the name, the version and the install folder from step 2, then approve what Claude Code asks to run. Expect REA to report no decompiler on Windows. That is the default there, and it was the state of Jason's PC: REA read two metadata files, and every other finding came from imports, exports and strings.
04
Read REPORT.md. Start with three sections: what REA could and could not see, the open-source projects the app is built on, and what could not be determined. The report is the outcome of this guide. Steps 5 and 6 are optional.
Your part: decide whether to go on. Go on when the report names an open-source project behind the main feature and finds no license check, account or server in the way. Jason's app was that case: its engine was a port of UxPlay, an open-source AirPlay server, and his PC already had a C++ compiler. Stop here when the report is thin. A thin report means the app's core is proprietary and REA had no decompiler, so there is no source to build from and no code to read.
05
Optional: paste the build prompt. It has Claude Code build your own version of one feature as a new program, from new code and open-source parts, in a new subfolder. Claude Code checks for a compiler before it starts, installs nothing system-wide, runs its own tests, and ends with a report on what it verified and what it did not.
Paste into Claude Code, with the four blanks filled in
Build my own version of one feature of the app described in REPORT.md. Read REPORT.md in the current folder first.
THE TARGET
- Original app: <app name>, version <version>
- Install folder: <full path of the install folder or app bundle>
- The feature I want: <the main feature, in one sentence>
I own this copy of the original app and I am allowed to examine it.
WHAT "MY VERSION" MEANS
- A new program that you build from source code. It runs on its own. It does not repack, load or depend on the original app's files.
- Put nothing from the original app in the result: none of its binaries, none of its artwork and none of its text.
- If REPORT.md shows that the original app is built on an open-source project, get that project from its own source repository and build it at a pinned commit. Write the commit in the README.
- Do not bypass or remove a license check, a sign-in or copy protection. If the feature depends on one, stop and tell me.
WHERE TO WORK
- Make a new subfolder named my-version in the current folder. Work only there.
- Never write into the original app's install folder. Never change the installed app.
- Use REA first when you need to look at the original app again. Open individual files: REA's open_binary tool accepts a file or a macOS app bundle and rejects any other folder.
BEFORE YOU WRITE CODE
1. Check that a compiler for this kind of program is installed. If it is missing, stop. Tell me what is missing and how to install it. Do not install it yourself. Build tools that can be downloaded into the my-version folder, such as CMake or Ninja, count as dependencies.
2. Install nothing system-wide. Download dependencies into the my-version folder only, and tell me how much disk space they use.
3. Tell me your plan in a few lines: what you will take from open source, what you will write, and what you will leave out.
WHILE YOU BUILD
- Build the feature I named first. Leave the other features out and list them.
- Each time you make a judgment call for me, add one line to ASSUMPTIONS.md: what you assumed, why, and what changes if you are wrong.
- When something fails, test one cause at a time. After three failed attempts at the same fault, stop and show me what you tried.
- Do not publish, push or upload anything. Ask me before any other action that sends something off this computer.
WHAT "DONE" MEANS
- Scripts that fetch the dependencies, build the program and package it, starting from an empty folder.
- A program folder that runs on its own on this computer.
- A log file that the program writes, and the path to it.
- A README: how to build the program, how to run it, and the license of every part.
- Self-tests that run without the real device or account. Tell me exactly which tests passed.
YOUR END REPORT, IN THIS ORDER
1. What you built.
2. How you built it.
3. What you verified, and how.
4. What you did not verify. Name the one test that only I can do, and give me the steps.
5. Guesses: each value or behavior you could not read from the original app.
6. What you left out.
7. The license of everything the program links, and the license of the result.
8. Firewall state: each rule or prompt the program needs.
9. Side effects on my desktop during your tests: windows, sounds, anything left running.
10. Disk space used.
11. What REA contributed, and what it could not do.
Your part: name the one feature you want, then watch the run. Jason used auto mode, and Claude Code asked him nothing across 193 tool calls. In the default permission mode, Claude Code asks you to approve most tool calls. His build took 47 minutes 2 seconds and left out recording, casting and the stats overlay. His first build also reused three FFmpeg library files that shipped with the original app. FFmpeg is open source, and a second build compiled it from source to replace them. This prompt rules that out from the start.
06
Test your version yourself, with the real device or account. Claude Code's self-tests run without it, so the end report names one test that only you can do. For Jason's app that test needed an iPhone: start the program, open Control Center on the phone, choose Screen Mirroring, and pick the PC's name.
Your part: run that test, then check every item the end report lists as not verified. If Windows asks whether to allow the program on your network, allow it on a private network, or the device cannot find the PC. If the test fails, tell Claude Code what you saw and point it at the program's log. In Jason's run the log shows video from his iPhone 53 minutes after the prompt. Sound from the phone was never tested. That test ran on his first build, and no device has connected to the second build.
To remove REA, ask Claude Code to run REA's uninstall command, npx rea-agents@latest uninstall. REA's docs say it "removes only REA-owned agent registrations and skill files". Your REPORT.md and your version stay in the working folder.
The specifics
From one run, on 12 October 2026.
Tested on
One Microsoft Store app on one Windows PC, with Claude Code v2.1.296 in auto mode and REA 6.4.0. Mac is untested.
What Jason's run did
Built in 47 minutes by the session record, with iPhone video at 53 minutes by the program's log. Sound from the phone is untested.
With no decompiler
REA made 8 of the run's 193 tool calls and read two metadata files. Every other finding came from imports, exports and strings.
Where your files go
REA's FAQ: "REA analyzes targets locally. Your agent receives the tool results, and its model provider has its own data policy."
What it is for
REA's README: "analysis of software you own or are authorized to examine". The build prompt forbids the original's code and artwork, and no prompt unlocks a paid app.
License of the result
Jason's version links the open-source UxPlay, so it is GPL-3.0: anyone who hands the program to someone else owes them the source.
Questions people ask
Will this work on any app?
No. It worked for Jason because the app's engine was a port of an open-source project, with no license check, account or server side, and his PC already had a compiler. With a proprietary core and no decompiler, the report comes back thin.
Is this legal?
REA's disclaimer says reverse engineering may be restricted where you live or by the license of the software you analyze. The prompts read an app you have installed, and the build prompt forbids the original's files in the result.
Will it change the app I already have?
The prompts tell Claude Code to leave it alone: read the app's files where they are, write only in a new working folder, and install nothing system-wide. In the default permission mode, Claude Code also asks you before most tool calls.
Can I share what it builds?
Read the license list in the end report first. Jason's version links UxPlay, so it is GPL-3.0, and passing it on means passing on the source. REA's disclaimer adds that an analysis report grants no rights to the original program.
Stuck on a step? Ask me.
This guide is also a lesson in my community, with its own questions thread for the step that stopped you. Joining is free. Paid members also get my ask me anything call every Tuesday at noon ET, and the full recordings of each setup as I make them.
Everything above this line is free and stays free. Below it is what we sell. The prices are the prices, and the two subscriptions carry 30 days money back on the first month.
Mica Social Media Analyzer
Your Instagram, answered by the ChatGPT or Claude you already use. Connect the account once, then ask in plain words what worked and what to post next.
Who it is for
You post on Instagram and you are done guessing from the Insights tab.
Price
$29/ month
What you get
Which Reels held attention, ranked, without opening Insights
Readings at 1, 24 and 72 hours after every post, so you see what happened while it was moving
Next captions written in the style of your most saved posts
Read only. It never posts or changes anything
The facts
7 days free with a card on file. About a dollar a day after that.
Our guarantee
30 days money back on the first month. Email guru@graniteai.co and we refund it.
Send us access to your repo. We run the nine-lane review, change nothing, and hand back the report with a 30 minute call on the three gaps to close first.
Who it is for
You have a product you charge for, or want to, and you would rather read the verdict than run the agents.
Price
$349One week
What you get
The report at docs/commercial-readiness.md: a verdict, a scorecard, every gap with the file to open and how to verify it
Legal, security, billing, support, marketing, operations and the rest, checked at once
A 30 minute call on the three gaps to close first
Nothing in your repo changed. The workflow prompt is yours to run afterward
The facts
The review is the one in the video. On our own product it found 87 gaps. The overnight run that followed was 148 commits and one pull request.
Our guarantee
If the review finds no gap worth your time, you pay nothing.
The first run, with Jason
We run the review on your repo together, on a call, and you leave with the report in your repo and the three gaps to close first.
Who it is for
You have the kit and a repo, and you would rather do the first run with someone who has done it.
Price
$9960 minutes
What you get
The review running on your machine, in your Claude Code
The verdict read together, and the three gaps picked
The workflow prompt set up so you can run it that night
The call recording
The facts
New. You would be one of the first. The run itself is the one in the video: 190 agents, 148 commits, one pull request.
Our guarantee
If the review will not run on your setup during the call, there is no charge.
Mica Social
An AI that writes, designs, checks and publishes a post to your Facebook and Instagram every day. You send a photo when you have one. It does the rest.
Who it is for
You know the feed should be alive every day and it is not.
Price
$49/ month
What you get
A finished post every day, with a photo made for it. Nothing to shoot
Timed to your local weather, season and county numbers
Five checks before anything publishes. No approvals, no calendar, no logins
Every photo you send becomes its own extra post
The facts
An agency running a daily feed bills $500 to $1,500 a month. 3 days free with a card on file.
Our guarantee
30 days money back on the first month. Email guru@graniteai.co and we refund it.