Jakeschincariol/replica-skill

Eleven Claude skills that walk a linear pipeline from reverse-engineering a target app through to deploying your own version on a custom domain. The chain is: recon, then architecture, design, build, backend, test, diff, entrepreneur, brand, launch, deploy — each step writes into a replica/ folder that the next step reads. You can run any skill in isolation, but the intended path is sequential.

The most interesting skill is /replica-entrepreneur, which reads a CSV of real user reviews from the App Store, G2, Capterra, and Reddit, then ranks complaints by frequency and links every quote back to its source row. The README is explicit that themes with fewer than three reviews or only one source get flagged as thin, and that no review without a URL is printed. That is a meaningful constraint — it forces the output to be auditable rather than plausible-sounding.

The six Python tools are all standard-library only, nothing to install. imgdiff.py is worth noting: it converts screenshots to edge maps before comparing them, so a color change in your clone doesn't register as a layout divergence. parity.py weights features as must/should/could, counts partials as half, and refuses to score intentional omissions against you. sweep.py searches for the original app's name, domain, and colors inside identifiers — not just string literals — and blocks the deploy step until the codebase is clean.

The legal framing is careful and specific. The skills are scoped to public pages and your own account. They won't read behind a login you don't own, won't touch private APIs, and won't copy proprietary assets. /replica-deploy won't ship until the brand sweep passes. The README explicitly says that if your account's terms of service forbid using it to build a competitor, the skill will say so and fall back to public sources only. Whether that self-enforcement holds in practice depends entirely on how the SKILL.md files are written — the README doesn't show those files directly.

The scheduling-app walkthrough in the README is the clearest part of the documentation. It shows /replica-recon producing 18 screens and 7 flows, /replica-diff scoring features at 86 and layout at 91, and /replica-entrepreneur surfacing 140 reviews with specific complaints about per-seat pricing and time zone confusion. Those numbers make the workflow concrete in a way that abstract descriptions of "clone any app" do not.

The honest caveat is in the fine print: a booking tool is weeks of work, a spreadsheet engine is not. /replica-recon is supposed to size the effort before you commit. That's the right place to put the expectation-setting, and the README doesn't oversell what the skills can do with genuinely complex applications.