Replica Skill
Eleven free Claude skills that reverse-engineer an app, rebuild its features, test it, and rebrand it as your own.
- Source repo
- Jakeschincariol/replica-skill
- Stars
- ★ 1.4k
- Last updated
- 6d ago
- License
- MIT
- Primary language
- Python
- FA score
- 53/100 · Major gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude CodeClaude.ai (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- An indie hacker sees that users of a scheduling tool hate per-seat price jumps: run recon through entrepreneur to get the feature map plus a positioning angle drawn from real review complaints.
- Not a fit if
- Teams that don't use Claude or Claude Code
- People who want to copy the target's code, assets or content
- Anyone expecting a one-shot clone of a complex app like a spreadsheet engine
- Source review
- 53/100 · Major gaps
What does this agent do, and when should you use it?
Replica Skill is a set of eleven Claude and Claude Code skills — recon, architect, design, build, backend, test, diff, entrepreneur, brand, launch and deploy — designed to be run in sequence from reverse-engineering through to going live. Each skill is a SKILL.md instruction file, and they pass work to one another through a replica/ folder in your project: recon writes screens, flows, components, an inferred data model and features.csv; architect plans the stack and schema; design measures screenshots into tokens; build reconstructs screen by screen; backend adds auth, database and payments; test produces Playwright specs; diff scores parity against the original. Six standard-library Python tools ship alongside: imgdiff.py for colour-agnostic layout diffs, parity.py for a weighted must/should/could parity score, reviews.py for ranking real user complaints with linked quotes, sweep.py for finding leftover names, domains and colours in your code, and listing.py for App Store and Google Play listing checks. The README states the skills rebuild functionality and UX patterns clean-room style and never copy source code, assets, logos, trademarks, copy, content or private APIs, and that deploy is blocked until the rebrand sweep is clean. Install by pasting the repo URL to Claude with "install skill", or as a Claude Code plugin from the marketplace. The tools need Python 3.8 or newer, install nothing via pip, and never touch the network.
The pipeline runs recon → architect → design → build → backend → test → diff → entrepreneur → brand → launch → deploy, with each skill reading what the previous one wrote into replica/. replica-recon reads the target's help center, pricing page, store listing and public walkthroughs plus your own account screenshots, and emits replica/recon.md (screens, flows, components, inferred data model) and features.csv. replica-architect chooses the stack and writes the database schema and API plan into architecture.md. replica-design measures screenshots into tokens.json — colour roles, type scale, 8px spacing — and contrast.py checks WCAG contrast on those tokens. replica-build builds the shell, then the flows, then every screen with its states, ticking off features.csv. replica-backend wires up sign-up, database, payments and integrations such as Google Calendar and Stripe. replica-test derives a test plan from the recon flows and writes Playwright specs for happy paths and edge cases, logging bugs by severity until no S1 or S2 remains open. replica-diff runs parity.py, which weights must/should/could, counts partial as half, ignores deliberately omitted and added features, and reports missing must-haves as not shippable; imgdiff.py converts both screenshots to edge maps and reports differing regions in original pixels so your new colours don't count against you. replica-entrepreneur drops every review without a link, clusters themes with counts and linked quotes, and marks themes thin when they have under three reviews or a single source. replica-brand generates a name, palette, logo brief and voice, then sweep.py scans the codebase — including identifiers like OriginalAppEmbed — for the original's name, domain and colours and blocks deploy while anything remains. replica-launch writes the landing page, pricing and store listing, and listing.py checks App Store and Google Play limits, ranking claims and copycat wording. replica-deploy runs preflight (tests, parity, sweep, listing), sets up production, returns DNS records for your domain and ships on your go.
- An indie hacker sees that users of a scheduling tool hate per-seat price jumps: run recon through entrepreneur to get the feature map plus a positioning angle drawn from real review complaints.
- A product manager scoping a build-versus-buy decision runs only /replica-recon to get an honest size estimate and full feature inventory before committing a team.
- A frontend engineer bootstrapping a new design system runs /replica-design to extract colour roles, type scale, spacing and component specs into tokens.json, then validates them with contrast.py.
- A QA engineer needs end-to-end coverage for a web app and uses /replica-test to generate Playwright specs from the recon flows, including time zones, double submit and a slot taken mid-booking.
- A founder doing pre-launch legal hygiene runs sweep.py for leftover names, domains and colours in the code, then listing.py to check store copy for limits and trademark issues.
- Anyone iterating on a clone in Claude Code uses parity.py and imgdiff.py after each pass to quantify remaining feature and layout gaps instead of eyeballing progress.
How do you install or deploy this agent?
Option one — paste this into Claude:
https://github.com/Jakeschincariol/replica-skill
install skillOption two — install as a plugin in Claude Code:
/plugin marketplace add Jakeschincariol/replica-skill
/plugin install replica-skill@replica-skillInstalled as a plugin, Claude Code namespaces the skills, so they show up as /replica-skill:replica-recon and so on.
Option three — copy the folders if you want plain names without the namespace:
git clone https://github.com/Jakeschincariol/replica-skill.git
cp -r replica-skill/replica-* ~/.claude/skills/For a project-local install instead of global, copy the same folders into your repo's .claude/skills/.
With no Claude Code at all, paste any single SKILL.md at the top of a chat and it runs as a mode; you lose the Python tools but the method works.
The tools need Python 3.8 or newer. There is nothing to pip install.
How do you use this agent?
Run the skills in order; each one reads what the last one wrote in a replica/ folder inside your project:
recon -> architect -> design -> build -> backend -> test -> diff -> entrepreneur -> brand -> launch -> deployA first invocation looks like this:
/replica-recon https://example-scheduling-app.comIt writes replica/recon.md and replica/features.csv, which the later skills consume.
The six Python tools run standalone:
python3 replica-diff/imgdiff.py original.png clone.png --out diff.png # layout diff, ignores colour
python3 replica-diff/parity.py replica/features.csv # parity score + missing list
python3 replica-entrepreneur/reviews.py replica/reviews.csv # what users hate, ranked, linked
python3 replica-design/contrast.py replica/design/tokens.json # WCAG contrast on your tokens
python3 replica-brand/sweep.py . --avoid "Original App" # anything of the original left?
python3 replica-launch/listing.py replica/launch/listing.json # store limits + copycat checksRun the bundled tests:
python3 -m unittest discover -s tests -vAny single skill can also run on its own — running /replica-entrepreneur on an app you are merely considering is a way to find out whether it is worth cloning.
What are this agent's strengths and limitations?
- Eleven skills form one ordered pipeline with all intermediate artefacts in a single replica/ folder, which makes staged review and iteration straightforward.
- All six Python tools use only the standard library and never touch the network, so they are easy to audit and need no pip install.
- parity.py has explicit scoring rules: must/should/could weighted, partial counted as half, deliberate omissions and additions excluded, and missing must-haves flagged as not shippable.
- reviews.py discards any review without a link, quotes only rows you supplied with that row's URL, and marks themes thin when they have under three reviews or a single source.
- sweep.py catches the original's name even inside identifiers and blocks the deploy until the sweep is clean, turning rebrand compliance into a hard gate.
- The README draws clear boundaries: functionality and UX patterns only, clean-room style, no copying of source code, assets, logos, trademarks, copy, content or private APIs.
- The core skills run on Claude / Claude Code; outside that environment you can only paste SKILL.md as a prompt and lose all Python tooling.
- Several steps depend on third-party or paid services (Stripe, Resend, Google Calendar, Playwright) whose keys and accounts you must supply yourself.
- No guarantee of a perfect clone: complex apps such as a spreadsheet engine are explicitly out of reach, and recon only sizes the work rather than promising it.
- Thin ecosystem evidence beyond the repo: no CI, release history or hosted offering is documented, and the bundled tests cover the Python tools only.
- Output quality depends heavily on inputs you prepare — reviews.csv, features.csv and screenshots are not fetched automatically.
- Legal risk stays with you: the README states this is not legal advice and tells you to run trademark checks and talk to a lawyer before selling.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Replica Skill This agent | 53 · Major gaps | Agent plugin / skillFree + model costs | ★ 1.4k | 6d ago | Python | Claude Code |
| Webwright Browser Coding Agent | 66 · Some gaps | CLIFree + model costs | ★ 6k | 3d ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| Auto RE Agent | 61 · Some gaps | CLIFree + model costs | ★ 2k | 1mo ago | Python | Codex · Claude Code · OpenAI API · Claude API |
| Cheat Engine MCP Bridge | 59 · Major gaps | MCP serverFree + model costs | ★ 1.6k | 1mo ago | Lua | Codex · Claude.ai |
How does FollowAgents rate this agent?
Why each dimension lost points
README claims the tools are pure standard library and never touch the network, and states it only reads public pages and the user's own account, never logging in elsewhere or bypassing paywalls; sweep.py blocks deploy on residual branding. Those are positives. But least_privilege is only a prose promise with no code-level constraint on filesystem writes or network access; user_confirmation appears only at deploy ('ships it when you say go') while the other ten skills (including backend wiring Stripe/Google keys and deploy) lack stepwise confirmation; data_flow_transparency does not say how recon data is stored or flows; sensitive_data_handling never addresses key or credential storage/redaction; external_effects involve real deployment, payments and third-party APIs with no sandbox or blast-radius note; rollback has no undo path at all; source_attribution scores 2 for clear MIT license and author credit. Deductions reflect many promises and little code-level evidence.
README and the tests directory are consistent: six tools each have a test file covering PNG filters/bit depths, contrast thresholds, listing length and banned words, so self_consistency is 2. dependency_availability is 2 because it declares Python 3.8+ standard library only with nothing to pip install. failure_messages is acceptable: PngError, CLI exit codes 0/1/2, and an interlace refusal with advice. No run logs or CI config are provided, so nothing reaches 3.
audience_and_scenarios is 2: it targets developers who want to clone and commercialize an app, with a full 11-step scheduling-app walkthrough. capability_boundaries is 2: it explicitly limits itself to features and flows, not code/assets/network/licences, and warns complex apps will not clone perfectly. trigger_precision is 1: only slash-command names and prose, with no trigger conditions, input validation or ambiguity handling. environment_fit is 2: it covers Claude plugin, copying into ~/.claude/skills, project-local .claude/skills, and pasting a single SKILL.md.
information_architecture is 2: a table of the eleven skills plus a file tree. install_notes is 2: plugin and manual install commands. naming_stability is 2: consistent replica-* naming. examples_and_faq is 2: a full scheduling-app walkthrough and tool usage examples. known_limitations is 2: the fine print covers no perfect clone and required trademark/legal checks. license is 3: full MIT text. versioning_changelog is 0: entirely absent. maintenance_responsibility is 1: only an author and site credit, with no issue channel, support commitment or update cadence.
output_usability is 2: each skill produces named artifacts (recon.md, features.csv, tokens.json, test-plan.md) and tools emit scores and missing lists. marginal_value is 2: an end-to-end recon-to-deploy pipeline adds value over single prompts. cost_benefit is 2: free, MIT, no API key, standard library only. None reach 3 because no actual output samples or outcome evidence are provided.
claim_traceability is 1: reviews.py requires a link per quote and a stated sample size, which is traceable by design, but only at the design level with no real data. cross_source_corroboration is 1: test files and README descriptions corroborate each other, but there is no independent third-party verification. fact_inference_separation is 1: recon distinguishes an 'inferred data model' from public facts and the Entrepreneur claims no fabrication. Scores stay conservative for lack of verifiable run evidence.
- README promises 'never touches the network' and 'only reads public pages', but no code-level network or filesystem constraint is visible; static review cannot confirm the claim.
- Apart from deploy, the other skills (including backend wiring Stripe/Google keys and going live) lack stepwise user confirmation, creating misoperation risk.
- It does not describe storage, redaction or flow of recon data, API keys or account credentials; sensitive-data evidence is thin.
- There is no rollback or undo path, so external effects such as deployment and payments are hard to reverse.
- No version number or changelog, and no issue channel or maintenance commitment, so long-term maintainability is unknown.
- Publisher identity is unverified by the FollowAgents enterprise registry; unknown rather than suspicious, and must not be used to infer reliability.
- This is a static source review; no code was executed, so test passing and actual tool behavior are unverified.