Someone Built Eleven Claude Skills That Clone Any App. People Are Already Cloning Those.

Replica is eleven free Claude skills that take an app apart and rebuild it. Recon maps the screens and flows. Architect picks the stack. Build writes it, test clicks through it, then diff scores your copy against the original. One called Entrepreneur reads real user reviews of the app and turns the complaints into your positioning. MIT, 18 days old, 3,973 stars. The fork ratio is 17.2%, the highest on anything we tracked this month, so people are running it rather than saving it. The guardrails are built into the skills themselves, including a rebrand sweep that blocks deploy until nothing of the original is left. We also found at least four accounts that re-uploaded the repo word for word under their own names.

Best for anyone studying how a product is built. Not ideal for anyone expecting a legal shortcut.

Every few months somebody ships a tool that makes people uncomfortable. This one arrived quietly, then went straight to work.

Eleven skills that promise to clone any app sounds like the kind of claim that stays a claim. Here, the numbers say otherwise.

What Replica Actually Is

Eleven skills for Claude Code. Free and MIT licensed, with no signup and no API key beyond whatever you already pay for. You install them and get eleven slash commands that run in sequence.

Above all, the pipeline is the design. Each skill reads what the one before it wrote, into a replica/ folder in your project, so the output of recon becomes the input of architect and so on down the line.

Here’s the repo, verified off the GitHub API in this morning’s scan.

As a result, installation takes one of four shapes. Paste the repo URL into Claude with “install skill” and it handles it. Add it as a plugin in Claude Code with /plugin marketplace add. Clone it and copy the folders into ~/.claude/skills/. Or drop them into your project’s own .claude/skills/ instead.

Installed as a plugin the commands get namespaced, so you type /replica-skill:replica-recon. Copy the folders manually and you get the plain names. The tools want Python 3.8 or newer and there’s nothing to pip install.

No Claude Code at all? Paste a single SKILL.md at the top of a chat and it runs as a mode. You lose the Python tools and keep the method.

The Fork Ratio Is The Story

Stars are cheap. In fact 3,973 of them on an 18 day old repo tells you it got shared, nothing more than that.

685 forks against those stars is 17.2%. That’s the number worth stopping on. A fork means somebody took a copy to their own account, which is one more step than a star and one step closer to actually using the thing.

Here’s where that sits against everything else we logged this month.

Under 12% reads as a bookmark pile. Google’s harness2 pulled four times the stars and sits at 9.3%, which means most of those eleven thousand people saved it for a weekend that never came.

Replica is at the top of the board. Whatever people think of it, they’re taking it home.

Why forks specifically? Because this thing is a folder of skill files rather than an app. To use it at all you either clone it or copy the folders, so forking is closer to installing than it is on a normal repo. That inflates the ratio compared to something you would pip install, which is worth saying before anybody treats 17.2% as a like-for-like number against a Python package.

Even so, the comparison above holds, because harness2, DigUp and coucou all ship code you copy too. Among repos of the same shape, this one is the one people are taking.

One caveat on all of those. A ratio on a young repo is noisy, since stars arrive faster than forks in the first week. Replica is 18 days old now, which is long enough for the number to mean something and not long enough to be settled.

The Skill Nobody Expected To Be The Good One

Meanwhile, ten of these skills do roughly what the names say. One does something else. It’s the one worth copying even if you never clone anything.

/replica-entrepreneur reads real reviews of the app you’re studying. The README’s worked example of cloning a scheduling tool has it reading 140 reviews pulled from the App Store, from G2 and Capterra, then from Reddit on top.

What comes back is a ranked list of complaints with counts and linked quotes attached. In that example the top three were per-seat price jumps, missing text reminders, then guests getting confused by time zones. Then it writes a fix plan: flat pricing, SMS reminders, a time zone confirm step. Then a positioning angle, which in the example was booking links for small teams that hate per-seat pricing.

The fine print on this one is firm. The Entrepreneur “never makes things up,” no invented reviews or quotes or numbers. Every claim carries a link, with the sample size stated.

Strip the cloning away and that’s competitive research, done properly, in an afternoon.

Think about how that work usually happens. Somebody skims a dozen reviews, forms an impression, then argues for it in a meeting on the strength of a feeling. A ranked list with counts behind it settles that argument before it starts, because you can point at how many people said the same thing. Most people building anything have never read 140 reviews of the thing they’re competing with. The skill is free and it works on any product with public reviews, including yours.

If you’re building something to sell, our piece on AI digital products and passive income covers the part that comes after the research. This covers the part most people skip first.

The Guardrails Are In The Tool

This is where it gets more interesting than the pitch suggests. The limits are written into the skills themselves rather than into a disclaimer nobody reads.

Six of them, from the README’s own fine print.

It uses public sources and your own account. Nothing it does gets past a paywall. If your account’s terms forbid using the product to build a competitor, the skill says so and sticks to public material.

It always rebrands before launch. New name, new palette, new logo, new words. /replica-brand runs a sweep through the codebase and /replica-deploy refuses to ship until that sweep comes back clean.

“Clone any app” means the features and the flow. That rules out the content, the network, plus every licence a company owns. The README’s example is blunt about it. You can rebuild a music app’s player, its playlists and its sharing. You can’t clone its catalogue. In the scheduling example the partner marketplace gets marked out of scope, on the grounds that a network isn’t a feature.

There’s no promise of a perfect copy. A booking tool is weeks of work. A spreadsheet engine isn’t happening at all. /replica-recon is supposed to size that up front rather than after.

/replica-diff gives you a number instead of a feeling. The worked example scores features at 86 and the booking page layout at 91, with round-robin scheduling and the website embed listed as missing.

Lastly it says to run the trademark checks, read the terms of everything you used, then talk to a lawyer before launch if there’s money involved. That section adds that it isn’t legal advice, which is correct and also the only line in the file that sounds like somebody imagining a courtroom.

The Part We Did Not Expect

While checking what coverage this repo already had, we found something better than coverage.

At least four other GitHub accounts have re-uploaded replica-skill under their own names, carrying the identical README down to the phrasing. One of them renamed it clone-replica-skill, which at least has the virtue of honesty.

So the tool for cloning things is being cloned, by people who didn’t bother to use it. There’s no recon step here and no rebrand sweep. They copied the files and changed the owner name.

Because it’s MIT licensed, redistributing it is allowed. MIT asks for one thing in return, which is that the copyright notice travels with the code. We’re not in a position to audit whether each of those copies kept it.

Still, what it demonstrates is the gap the original is actually trying to close. Replica’s whole method is the opposite of what those accounts did: study the thing, rebuild it yourself, make it different, put your own name on it. The lazy version takes four seconds and produces nothing.

Is It Okay To Clone Any App

Worth addressing directly, since it’s the first question anybody sensible asks.

The Precedent

Rebuilding a competitor’s feature set from public behaviour is old and it isn’t, by itself, illegal. Compaq reverse engineered IBM’s BIOS in 1982 with a clean room process and built an industry on it. Clean room means one team studies the original and writes a description. A second team then builds from that description without ever seeing the source. The point is to show the copy came from observed behaviour instead of from copied code.

Replica is doing a version of that. Recon works from public pages, store listings, walkthroughs and your own account, then build works from the recon map.

Courts have looked at this shape before. The cases that went badly for the copier generally involved copied code or copied content rather than copied behaviour, which is the distinction the clean room process exists to make visible. None of that is a promise about your situation, since the law here turns on specifics that a README can’t know.

The part that’s new is the compression. What took Compaq a team and a year now takes one person a weekend. On top of that the Entrepreneur skill bolts a market gap analysis onto the front, so the copy ships with a reason to exist. Nothing there changes the law. It changes the volume. Volume has a way of turning a legal activity into a problem anyway.

The Case For It

The other side of the argument is worth stating too. Every SaaS category already has fifteen near-identical products built by people who studied each other. Replica does openly and cheaply what agencies have charged for quietly. And the complaints it surfaces are real complaints from real users that the incumbent chose not to fix. Somebody fixing them is how markets are supposed to work.

Where it clearly goes wrong is the obvious place. Clone the catalogue. Lift the copy. Keep the branding close enough to confuse. Ignore the terms you agreed to. No skill file is going to protect you from any of that. The README says as much. Whether anyone reads that part is a separate question. The four re-uploads suggest an answer.

Verified vs Asserted

How It Works In Practice

The README walks a full example, cloning a scheduling app. It’s specific enough to be worth following.

Recon Through Build

Recon reads the help center, the pricing page, the store listing and public walkthroughs, while you click through your own account alongside it. Out comes a recon map: 18 screens, 7 flows, the components, an inferred data model covering users and event types and availability and bookings, plus a features.csv matrix.

Architect picks Next.js with Postgres, Stripe for payments and Resend for mail. It writes the schema with a constraint so two guests can never book the same slot, then orders the build so the booking flow goes end to end first.

Design measures the screenshots into tokens. Colour roles, a type scale, 8px spacing, specs for the date picker and the slot button. Open icons, an open font, your own words.

Build does the shell first, after that the booking flow, finally every screen with all of its states, ticking off features.csv as it goes. Backend adds signup, Google Calendar through Google’s own API with your keys, Stripe for paid bookings, reminder emails.

Test Through Deploy

Test writes a plan from the 7 flows and Playwright specs for both the happy paths and the edge cases. Time zones. A slot taken mid-booking. Double submit. It logs bugs by severity and keeps going until no S1 or S2 is open.

After that, diff scores it. Then entrepreneur reads the reviews, brand renames it, launch writes the landing page, then finally deploy ships it.

Notably, each skill ships with its own tooling rather than just a prompt. Design has a contrast.py. Diff has imgdiff.py and parity.py. Entrepreneur has reviews.py. Brand has sweep.py. There’s a tests/ folder covering every tool, which is more than most repos in this category bother with.

Our roundup of GitHub repos worth running with Claude Code covers a lot of single-purpose skills. An eleven stage pipeline with shared state between stages is a different kind of thing. It’s also the direction this category keeps moving.

Who This Is Actually For

Three groups. Only one of them is the obvious one.

Before any of that, a caveat about expectations. Although the pipeline runs end to end, every stage still hands you work to review. Recon infers a data model rather than reading one. Build writes code you have to check. Treat it as a very fast junior who read the whole product first, because that’s roughly the level it operates at.

People Studying A Product

Firstly, the best use costs you nothing and breaks no rules, because you never clone any app at all. Run recon on a product you admire and read the output. You get its screens, its flows and its inferred data model written down, which is a free education in how that category is built.

Stop there. You’ve learned more in an afternoon than from a month of reading its marketing site.

People Who Already Have A Competitor

Secondly, run the Entrepreneur skill on the incumbent in your market and you get back ranked complaints with counts and links. That’s your roadmap, written by their customers, for free.

For someone already building, this is the highest value thing in the repo. It requires cloning absolutely nothing.

People Who Want To Ship A Clone

You can, though the tool tries hard to keep you on the right side of the line. Follow the rebrand step. Honour the terms you agreed to. Leave the catalogue and the network alone, then run the trademark checks it hands you.

Then talk to a lawyer, because the README tells you to and because a parity score of 86 isn’t a legal opinion.

For what it’s worth, the people most likely to get burned here aren’t the ones building a careful clone. They’re the ones who skip straight to deploy, which the tool specifically refuses to let you do. Whether that gate survives somebody determined to remove it is a different question, since it’s MIT and the file is right there.

What It Costs

The skills are free and MIT licensed. Your bill is the Claude Code subscription or the API credits underneath them. An eleven stage pipeline across a real app isn’t a cheap afternoon of tokens.

Our Claude Pro review covers what the subscription tiers include. Running recon alone is light. Running build and test across 18 screens is where the usage goes. That’s before diff starts comparing screenshots.

Besides this one, the author ships four sibling skills built the same way. Those ones target Arena and X, plus LinkedIn and Instagram. Same pattern, different target. The skills scene has been consolidating into bundles like this for months, which we covered in the piece on the agent skills marketplace.

How Much Has Been Written About This

A note on how much has been written about this. Earlier this week we got a claim like this wrong somewhere else, so the fix is to show the work rather than assert it.

What we found: one blog writeup, plus listings on three Claude plugin directories. No post surfaced on Reddit in our search. No major outlet. Four verbatim re-uploads on GitHub, which are copies rather than coverage.

In short, that’s the extent of what we checked, so that’s the extent of what we’ll claim. It says nothing about forums we didn’t search or conversations we can’t see.

For a repo with 685 forks, that’s a thin paper trail. Make of it what you want.

The Part Worth Keeping

Eleven skills, 3,973 stars, 685 forks. The highest fork ratio on anything we tracked this month. Guardrails written into the tool rather than bolted on, which is rarer than it should be.

Meanwhile the clone pipeline is what gets all the attention. Review mining is what you should actually take away.

Read 140 reviews of whatever you’re competing with. Count the complaints, fix the top three, then say so on your landing page. That’s the entire method. It was always available, yet it took an app cloning tool to get anybody to automate it.

Charts and Blocks

Fork Ratio, This Month’s Repos

The Pipeline