reverse-engineer-and-rebuild
/reverse-engineer-and-rebuild
Odtwarza działającą stronę, aplikację lub API na podstawie adresu URL albo dostarczonego pliku i porównuje oryginał z kopią.
- Autor
- cevo1466 ↗
- Repozytorium
- Reverse-Engineering ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 31.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 18
- Skrypty
- 13
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 31.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 13 skryptów w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/capture_http.py, scripts/capture_scroll.py, scripts/capture_web.py, scripts/compare_capture.py, scripts/compare_http.py, scripts/compare_scroll.py, scripts/compare_web.py, scripts/extract_tokens.py, scripts/pixel_diff.py, scripts/rea_once.py, scripts/report_safety.py, scripts/verdict.py, scripts/viewport_matrix.py
Co robi i kiedy
Odtwarza działającą stronę, aplikację lub API na podstawie adresu URL albo dostarczonego pliku i porównuje oryginał z kopią.
Oryginalny opis (EN)
Rebuild a runnable website, app, API, or feature from a user-supplied URL or shipped artifact. Use when the user asks to reverse engineer and clone, recreate, port, or replace observable behavior. Gather evidence with REA or available tools, implement the rebuild, compare original and rebuild, and close observed gaps.
Jak działa
- Choose the observable contract
- Observe and trace
- Build, compare, refine
- Boundaries
Podgląd SKILL.md
--- name: reverse-engineer-and-rebuild description: Rebuild a runnable website, app, API, or feature from a user-supplied URL or shipped artifact. Use when the user asks to reverse engineer and clone, recreate, port, or replace observable behavior. Gather evidence with REA or available tools, implement the rebuild, compare original and rebuild, and close observed gaps. --- # Reverse engineer and rebuild Deliver a **working reconstruction**, not only an analysis report. One invocation should carry the work from target observation through implementation and verification. Work in the user's language and existing project unless they specify another destination. Aim for the exact observable result within the requested scope; never turn an unobserved feature into a claim of exactness. ## Choose the observable contract 1. Identify the target URL or artifact, requested output, existing workspace, and the routes, screens, or behaviors the user named. When no scope is named, start with the reachable primary flow and its visible states, then continue through other reachable routes and interactions. If no target is available, request it. 2. For a website or web app, read [web reconstruction](references/web.md). If the flow uses an API or persistent state, also read [backend reconstruction](references/backend.md). For an APK, binary, .NET assembly, Electron package, archive, or CLI, read [target routing](references/targets.md). These paths define what can actually be observed and compared. 3. Record a compact contract before coding: each route or input, action, expected visible/output state, viewport or runtime, evidence location, and confidence. Include loading, empty, error, and responsive states when they are visible or requested. Keep unknown behavior marked unknown. ## Observe and trace 4. Use REA when an existing CLI or MCP connection is available. The bundled first pass needs a new report directory: ```bash python3 <this-skill-directory>/scripts/rea_once.py TARGET --output REPORT_DIR ``` `REPORT_DIR` must not exist. The helper saves raw JSON, stderr, a manifest, and a readable report; it **does not build a clone**. Each REA command has a 600-second default limit, configurable with `--timeout-seconds`. It validates JSON, operation, and normalized result before marking a step successful. It can download pinned REA through `npx` when no CLI exists. Pass `--no-download` to use only installed tools. If REA is absent or unsuitable, follow [evidence fallback](references/fallback.md) and continue. Read relevant Evidence and its `limitations`, `coverage`, unknowns, and source identity. A successful command is not a completeness verdict. 5. Investigate the requested behavior beyond the first pass. For web, capture named user journeys at matching viewport/device settings with DOM, accessibility, screenshot, URL, and relevant events; use a page-specific ready condition. When REA is unavailable but Chrome or Chromium is installed, use `scripts/capture_web.py` for repeatable PNG, DOM, and accessibility captures; see [web reconstruction](references/web.md). For backend flows, record request, response, and state transitions with valid, invalid, and read-after-write cases; use `scripts/capture_http.py` for repeatable scenarios on authorized, resettable test instances. For packaged software, trace the relevant entry points and collect runtime cases where execution is available. Cite exact Evidence or capture paths in the contract and identify non-REA provenance. ## Build, compare, refine 6. Implement the runnable reconstruction in the user's workspace. Follow the project's stack and conventions; if it is empty, choose a maintainable minimal stack. Implement the content, assets, layout, navigation, responsive behavior, state transitions, and data/error behavior supported by the contract. For observed backend flows, implement actual routing, validation, persistence, error semantics, and frontend integration; follow [backend reconstruction](references/backend.md). Do not finish with a static mockup when the target responds to input. Keep independently created code and assets organized so the app builds and runs. 7. Run the rebuild. Execute each recorded journey against the original and rebuild under equivalent conditions. Capture the same states; compare visual layout, text and accessibility, URL/navigation, controls, state changes, and relevant network/process output. For web, use `scripts/compare_web.py` on two REA captures, `scripts/compare_capture.py` on two Chrome capture manifests, or `scripts/pixel_diff.py` on matching PNG screenshots from another tool. Compare backend transcripts with `scripts/compare_http.py`, including state after writes. For other targets use the applicable REA comparison or fixture workflow from [target routing](references/targets.md). Do not treat static similarity as runtime parity. To check several responsive layouts in one run, use `scripts/viewport_matrix.py` (default 375x812, 768x1024, 1440x900); it writes a combined verdict with viewport-prefixed case IDs. For scroll-driven or hover animation, record both sides with `scripts/capture_scroll.py` and compare them with `scripts/compare_scroll.py`; it checks settled states at the sampled scroll positions, not the easing between them. To seed the rebuild's colours, type scale, spacing and radii from the original, run `scripts/extract_tokens.py` on the target and treat `tokens.json` as a ranked starting point, not a design system. 8. Fix observed material differences, recapture the affected cases, and repeat until the requested journeys pass or a specific blocker makes further observation impossible. Keep an evidence-to-implementation checklist with `matched`, `different`, or `unknown` for each case. Never mark a case matched solely because code exists or tests pass without comparing the original. Combine the comparison outputs into that checklist with `python3 <this-skill-directory>/scripts/verdict.py RESULT_OR_DIR [RESULT_OR_DIR ...] --output VERDICT_DIR`; it writes `verdict.json` and `verdict.md` and exits 0 (all matched), 1 (any different) or 2 (unknown or error). 9. Deliver the runnable files, the exact local run command, verification evidence, and a concise list of remaining differences or unknowns. State the scope of any “birebir” or exactness claim: the tested routes, states, inputs, viewports, and environment. If a prerequisite prevents the build or comparison, still finish all independent implementation and checks, then identify the precise blocked cases. ## Boundaries - The first-pass helper opens a supplied web URL but does not execute local target programs. Running an original binary, attaching to a process, or using a live account can affect external state; use the user's authorization and environment constraints for those observations. - Do not silently install native analysis engines, bypass access controls, or publish captured target data. Do not replay writes against production or a user's live account without authorization and a way to isolate their effects. REA may be downloaded by `npx` when the helper cannot find an installed CLI; `--no-download` prevents that. - All bundled helpers write a local `.gitignore` inside each new output directory, regardless of its name. Keep raw target observations out of commits and review them before sharing. - REA is maintained by [morluto/rea](https://github.com/morluto/rea). This independent skill orchestrates its public interfaces and does not bundle REA.
Pliki w paczce (18)
- SKILL.md 7 KB
- references/backend.md 7 KB
- references/fallback.md 2 KB
- references/targets.md 3 KB
- references/web.md 8 KB
- scripts/capture_http.py 16 KB
- scripts/capture_scroll.py 9 KB
- scripts/capture_web.py 20 KB
- scripts/compare_capture.py 6 KB
- scripts/compare_http.py 11 KB
- scripts/compare_scroll.py 9 KB
- scripts/compare_web.py 14 KB
- scripts/extract_tokens.py 12 KB
- scripts/pixel_diff.py 11 KB
- scripts/rea_once.py 19 KB
- scripts/report_safety.py 1 KB
- scripts/verdict.py 9 KB
- scripts/viewport_matrix.py 7 KB
Wersje
W bazie nie ma innych skilli o nazwie reverse-engineer-and-rebuild.
Migawka bazy z 13.11.2026 · wersja paczki 582aaa5ffd94 (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100clone-website
Odtwarza stronę internetową, wydobywając sekcja po sekcji zasoby, CSS i treść oraz uruchamiając równoległych agentów budujących w worktrees. - 100clone-site
Klonuje dowolną stronę z adresu URL do wiernej kopii w Next.js: rozpoznanie, wyciągnięcie zasobów i CSS, równoległe budowanie sekcji, port animacji i automatyczna weryfikacja zgodności. - 100rea
Ustala, jak dokładnie działa funkcja w gotowej aplikacji bez kodu źródłowego, dobierając narzędzie analizy i zostawiając dowody poza kontekstem. - 100recreate-recon
Tworzy mapę ekranów, przepływów, komponentów i modelu danych aplikacji z publicznych stron, zrzutów i opisów w sklepach, jako pierwszy krok. - 100clone-website
Odtwarza stronę internetową 1:1 na podstawie pomiarów w Figma, Penpot lub kodzie i weryfikuje wynik względem oryginału. - 100phoenix
Przebudowuje istniejącą aplikację w wersję drugą etapami: mapuje ją, planuje architekturę, buduje, testuje i porównuje z oryginałem.