Skip to content

About

Four Claude skills that rebuild an app and jump past it with what its users already asked for.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

Leapfrog

Leapfrog: rebuild the app, then jump past it

Four Claude skills that rebuild an app and then jump past it.

Most clones copy the features and stop there. Leapfrog also mines what the original's users already asked for (feature requests, workarounds, the reasons they left) and builds those in, so your version has a reason to exist.

Pipeline: scout, build, verify, ship. User wishes found in scout become the leaps you ship.

skill what it does
/leapfrog-scout Maps screens, flows and data model. Mines user wishes and competitor gaps. Picks the leaps and the angle.
/leapfrog-build Stack, SQL schema, design tokens from computed CSS, screens with every state, auth, payments, integrations.
/leapfrog-verify Playwright tests from the flows, bug log, feature parity, leap progress, layout diff that ignores colour.
/leapfrog-ship New name and brand, leftover sweep, landing page led by the leaps, pricing, store listing lint, preflight, deploy.

Free, MIT, no API key, nothing to connect.

Built to be cheap on tokens

The skills keep Claude's context small. Scripts do the bulk reading and scoring and print a short summary; Claude opens a full report only when it needs one.

Descriptions 5,400 to 932 chars, skill text 48 KB to 17.6 KB, review mining about 100k to about 2k tokens (estimate)

work how Leapfrog does it
finding the app's pages probe.py reads sitemaps, groups URLs (help, pricing, changelog, wish boards...). No browsing.
reading a page probe.py --text URL: plain text, no browser render.
collecting feedback wish_fetch.py pulls official feeds (App Store RSS, Hacker News, GitHub issues). Boards and forums go to a subagent that returns only matching rows.
ranking wishes wish_rank.py keeps only sentences that ask, work around or churn, clusters and scores them. Claude reads the top clusters, not 500 reviews.
design tokens extract_tokens.mjs reads computed CSS. No screenshots to measure.
schema model_to_sql.py drafts the SQL from the data model. Claude only reviews it.
layout checks Screenshots go to disk. layout_diff.py batch scores every screen in one call; Claude opens an image only when its score is low. --since re-checks only changed screens.
resuming leapfrog/state.json holds the step and the facts, so a new session reads one small file.
recon Sharded: an index plus one file per screen and flow. A build step loads one shard.

Install

As a plugin in Claude Code:

/plugin marketplace add Bambeek/leapfrog-skill
/plugin install leapfrog-skill@leapfrog-skill

Or copy the folders (keep leapfrog-shared, the skills use it):

git clone https://github.com/Bambeek/leapfrog-skill.git
cp -r leapfrog-skill/leapfrog-* ~/.claude/skills/

Python 3.8+ for the tools, standard library only. extract_tokens.mjs and the tests in your project use Playwright.

Tools

python3 leapfrog-shared/state.py show
python3 leapfrog-scout/probe.py https://app.example --out leapfrog/sources
python3 leapfrog-scout/wish_fetch.py --out leapfrog/wishes.csv --apple 123456789 --hn "App"
python3 leapfrog-scout/wish_rank.py leapfrog/wishes.csv --changelog leapfrog/sources/changelog.txt
python3 leapfrog-scout/opportunity.py draft leapfrog/wishes.json --gaps leapfrog/gaps.csv
python3 leapfrog-build/model_to_sql.py leapfrog/recon/model.md --out leapfrog/schema.sql
node    leapfrog-build/extract_tokens.mjs https://app.example/pricing --out leapfrog/design
python3 leapfrog-build/tokens.py all leapfrog/design/tokens.json --out src/styles
python3 leapfrog-verify/parity.py leapfrog/features.csv
python3 leapfrog-verify/layout_diff.py batch leapfrog/screens leapfrog/clone-screens
python3 leapfrog-ship/sweep.py . --config leapfrog/brand.json
python3 leapfrog-ship/listing.py leapfrog/launch/listing.json --brand leapfrog/brand.json
python3 leapfrog-ship/preflight.py

Tests: python3 -m unittest discover -s tests

Rules the skills follow

Full text in leapfrog-shared/RULES.md.

  • Rebuild what an app does, never what it owns. Features, flows and UX patterns, written fresh. Not its code, assets, logo, fonts, copy or private APIs.
  • Public sources and your own account only. Official feeds within their terms, everything else read like a person reads it. No scraping, no logging into other accounts, no paywall tricks.
  • No invented evidence. Every quote is verbatim and linked, sample sizes are stated, thin evidence is labelled thin. No fake reviews or testimonials.
  • You own the accounts and the money. Claude never signs up, pays or pastes live keys. Nothing goes live without your go.
  • Ship only as your own. New name, palette, logo and words. The deploy waits until the sweep is clean.

Name checks are screens, not legal clearance. Talk to a lawyer before you launch if money is on the line. This is not legal advice.

License

MIT.

About

Four Claude skills that rebuild an app and jump past it with what its users already asked for.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages