Replica, an eleven-skill kit for Claude that reverse-engineers an app and rebuilds it clean-room, hit 119 GitHub stars on day one

An indie maker who looks at a popular product and thinks they could do it better usually spends weeks mapping its screens, drafting a plan, and reading reviews before writing a line of code. Replica, released on 3 October 2026 by Jake Schincariol as an MIT project, packs that work into eleven Claude skills that chain into one run.

The repository is at github.com/Jakeschincariol/replica-skill and the GitHub REST API reports 119 stars and 10 forks on 4 October 2026. It is the fifth skill kit from Schincariol following similar bundles for LinkedIn, Arena, YouTube and X.

What the eleven skills do

replica-recon reads an app's public pages, pricing, store listing and public walkthroughs, and asks the user to click through their own account. It writes a map of screens, flows, components and an inferred data model into a replica/ folder in the project. replica-architect picks a stack and a database schema from that map, replica-design turns screenshots into design tokens, replica-build rebuilds screen by screen, and replica-backend wires up auth, payments and integrations.

replica-test writes a test plan from the mapped flows and files bugs by severity until no S1 or S2 is open. replica-diff compares the clone to the original and reports a parity score with the missing features. replica-entrepreneur reads real reviews on the App Store, G2, Capterra and Reddit, groups the top complaints with counts and linked quotes, and writes a fix plan. replica-brand, replica-launch and replica-deploy rename everything, write the landing page and ship to a domain.

The clean-room rule, enforced by one of the tools

Replica never copies source code, logos, trademarks, copy or private APIs: it studies what the app does and how a person moves through it, then writes everything fresh. The repository's own replica-brand/sweep.py scans the codebase for the original's name, domain and colours, including inside identifiers like OriginalAppEmbed, and replica-deploy refuses to ship until the sweep is clean. The project also says replica-recon only reads public pages and the user's own account, never scrapes behind a login, and reads only public pages if the user's account terms forbid using it to build a competitor.

Four of the six Python tools are standard-library only and make no network calls. imgdiff.py compares two screenshots using edge maps so the clone's new colours do not count against its layout score. parity.py weights must, should and could features, counts partial as half, and never scores items that were cut on purpose. reviews.py drops every review without a link and prints only quotes copied from rows the user provided, with each row's URL. listing.py checks App Store and Google Play limits, ranking claims in the title, and the original's name anywhere in the listing.

What the author says is still hard

The repository states that results depend on how complex the target app is: a scheduling tool is a few weeks of work, a spreadsheet engine or a networked game is far more. replica-recon sizes a job honestly before anyone starts, and replica-diff returns a measured parity number rather than a judgement. The author also warns to run the trademark checks replica-brand lists and to speak to a lawyer before launch if the clone will be sold.

The kit installs by pasting the GitHub URL and the words "install skill" into Claude, or through the Claude Code plugin marketplace, or by copying the folders directly into ~/.claude/skills/.

Source

• Replica GitHub repository, primary source: https://github.com/Jakeschincariol/replica-skill
• GitHub REST API for star and commit information, queried 4 October 2026: https://api.github.com/repos/Jakeschincariol/replica-skill