hk-code-quality-audit
/hk-code-quality-audit
Wyszukuje martwy kod, wklejone szablony i niepotrzebne zależności, i zapisuje wynik jako stronę HTML z oceną na 10.
- Autor
- deepklarity ↗
- Repozytorium
- harness-kit ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 10.11.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 1
- Skrypty
- 0
- Komentarze HTML
- 1
- Adresy URL w SKILL.md
- 1
Brak sygnałów. Repozytorium zmienione 10.11.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- SKILL.md zawiera komentarze HTML (1). Nie widać ich w podglądzie Markdown, a model je czyta.
Co robi i kiedy
Wyszukuje martwy kod, wklejone szablony i niepotrzebne zależności, i zapisuje wynik jako stronę HTML z oceną na 10.
Wyzwalacze z opisu
Argumenty
/hk-code-quality-audit [backend | frontend | a directory | one category name, default: whole repo]
Oryginalny opis (EN)
Find what nobody would miss and what reads as generated, and write it up as one HTML page scored out of 10: dead code, pasted templates, wrappers with one caller, scratch files, hardcoded hosts, unused packages. Triggers on: 'find slop', 'AI slop', 'hygiene check', 'dead code', 'dead endpoints', 'unused dependencies', 'backend slop', 'frontend slop', or /hk-code-quality-audit.
Jak działa
- Scope
- Steps
- Backend scope, added to the categories
- Frontend scope, added to the categories
- Closing step
- Example
Podgląd SKILL.md
--- name: hk-code-quality-audit description: "Find what nobody would miss and what reads as generated, and write it up as one HTML page scored out of 10: dead code, pasted templates, wrappers with one caller, scratch files, hardcoded hosts, unused packages. Triggers on: 'find slop', 'AI slop', 'hygiene check', 'dead code', 'dead endpoints', 'unused dependencies', 'backend slop', 'frontend slop', or /hk-code-quality-audit." argument-hint: "[backend | frontend | a directory | one category name, default: whole repo]" allowed-tools: Bash, Read, Grep, Glob, Write, Agent --- # Code quality audit Reads the code and writes up, with the lines that show it, what could be deleted and what reads as generated: dead code, duplicated logic, abstractions with one caller, files in the wrong place, hardcoded hosts, packages nothing imports. It changes nothing. Secrets and auth are `hk-security-audit`'s; this skill stops at a hardcoded localhost URL or port. ## Scope <skill_context> $ARGUMENTS </skill_context> Empty means the whole repo. `backend` means the API or worker service directory CLAUDE.md names, `frontend` the frontend source tree it names. All six categories in step 2 run on every scope. `backend` or `frontend` also adds that layer's list at the end of this file. A path means that path. A category name from step 2 means only that category, over the whole repo. If the scope holds no source files, say so in one chat line and stop. ## Steps **1. Map the layout.** Read the root `CLAUDE.md` and the `CLAUDE.md` of each project it lists. A thing those files call deliberate is not a finding. Note the dependency manifests, the folders to skip (`node_modules`, `.venv*`, `build`, `dist`, and any orchestration or worktree state folder CLAUDE.md names), and the env-var fallback pattern the frontend uses. Done when you can name each project's manifest and its skip list. **2. Scan the six categories.** Dispatch one subagent per category in one message, each given the scope, the category text below, and asked for candidates as `file:line`, what is wrong, and the evidence line. Haiku for files, hosts and dependencies, which are pattern matches; sonnet for dead code, duplication and abstraction, which need the caller graph. For a scope over roughly fifty files, the sonnet ones may fan out per top-level directory. The main conversation re-judges every kept finding in step 3. - **Files.** A Python script in the frontend tree, a test beside its source, a config at the wrong level; names like `temp_*`, `scratch_*`, `debug_*`, `old_*`, `backup_*`, one-off scripts (`populate_*`, `fix_*`, `quick_*`); logs, databases and build output tracked by git; tracked files that belong in `.gitignore`; `TODO`, `FIXME` and `HACK` comments older than six months by `git blame`. - **Dead code.** Unused imports; functions, routes, serializers, tasks, components and hooks with no caller; state, props, fields and handlers set and never read; commented-out blocks; tests for deleted code; feature flags nothing reads; management commands and scheduled tasks nothing invokes. - **Duplication.** The same logic in two files; one template pasted into several files with the names changed; the same validation or business rule written in two modules; the same fetch, formatting or empty-state logic in more than one component; two naming conventions in one directory; docs that describe behaviour the code no longer has. - **Abstraction.** A helper, wrapper or service class with one caller or none; a service that only forwards to the data layer; a custom hook with one caller that could be inline; props passed through two or more components that never read them; a mock or stub standing in for a subsystem on a production path; try/except or null checks around code that cannot fail that way; a catch that re-raises with nothing added; a comment that restates the line under it; a suppressed lint or type error with no reason beside it; a file over 500 lines doing several jobs. - **Dependencies.** Packages in a manifest that nothing imports; two packages for one job; dev packages in the runtime list. - **Hosts.** `localhost`, `127.0.0.1` and fixed ports outside `.env.example`, settings, test fixtures, and the env-var fallbacks whose default `.env.example` documents. Done when all six have reported, including the ones that came back empty. A category is empty when it was searched and nothing survived, never because the scope is a layer. **3. Verify every candidate by opening the file.** Before calling code dead, grep for the name as a string: dynamic imports, registries, URL maps, management commands, task names. Before calling a file misplaced, check the nearest `CLAUDE.md`. Fixtures, examples and templates are there on purpose. A finding is kept only when you can quote the lines and say who is confused by them. Done when every kept finding has its quoted lines and every dropped one has a reason. **4. Set the severity and derive the score.** Severity is 1 to 5, 5 worst, by what it costs the next reader: | Severity | Finding | |---|---| | 5 | a hardcoded host production code reads; a mock or stub on a production path; two implementations of one rule that already disagree | | 4 | a file in the wrong place; one template pasted into three or more files; dead state, endpoints or handlers in code people edit weekly; docs that describe behaviour the code no longer has | | 3 | a wrapper with one caller; two implementations of one rule that still agree; an unused dependency; a stale config | | 2 | a naming inconsistency; an oversized file; a guard around code that cannot fail; a comment that restates the line | | 1 | a suppressed lint with no reason; a docstring that names the function again; an old comment | Score 0 to 10 and the band for each set of findings: [../README.md](../README.md). Inside the band, go to the top when the findings sit in files nobody has touched for months, and to the bottom when a pattern repeats across files. **5. Build the page.** Read `docs/playbook/tone_and_taste.md` and `docs/skills/knowledge/hk-writeup/SKILL.md`, then copy `docs/skills/knowledge/hk-writeup/SHELL.html` to `$(git rev-parse --show-toplevel)/writeup/code-quality-<scope with slashes replaced by dashes>.html` (`repo` when the scope is the whole repo) and replace only its `<!-- CONTENT -->` line and its `<title>`. The shell's style block stays as it is, so this page looks like every other page under `writeup/`; a stylesheet of your own is the wrong page. In this order: 1. Kicker "Code quality audit", the scope as the title, chips for the commit and the file count scanned. The score is a sentence in the Overview, never a chip. 2. **Overview.** Two sentences: what the scoped code does, and the score out of 10 with the sentence that derives it; then the finding count per severity, which categories came back empty (so the reader knows they were checked); then the band table as a `.tbl` in `.tblwrap` so the reader sees which row the findings landed on. 3. **Findings**, one `section` per category, worst first inside it. Severity 3 and up is one `details.row` each, numbered F1 onward with the number in a `<span class="sn">`: `.row-name` is the severity in words, `.row-status` the title as a sentence; the body holds `file:start-end` in `.mono`, the literal four to fifteen lines in a `pre`, one sentence on what is wrong, one on what it costs the next reader, one on the fix. Severity 1 and 2 go in one `.tbl` per category: file, what, fix. 4. **Patterns seen more than once.** One sentence each, only for patterns that repeat across files. Omit the section if none. 5. Footer: read by subagents, verified by hand at the named commit; how many candidates were raised and how many dropped in step 3. **6. Report back** in chat: where the page went, the score, the count per severity, and the five fixes with the best effort-to-impact ratio. Done when that message is sent. ## Backend scope, added to the categories Routes registered in a URL map that nothing links to or calls; querysets, serializers, tasks and models with no caller; the same rule written in a view and again in a serializer; a service class that only forwards to the data layer; exception handling that catches and re-raises with no added information. ## Frontend scope, added to the categories Components exported but never rendered; new object, array or function values created in render and passed to memoised children or effect dependencies; inline styles that repeat a class already in the shared stylesheet; `useEffect` that mirrors props into state. ## Closing step Then the closing step in [../README.md](../README.md#closing-step): board rows for severity 4 and 5, a pattern for a recurring cause, one ledger line. ## Example - List the projects, for example `backend/`, `frontend/`, `cli/`. - Say what `backend` and `frontend` mean as paths, and where the shared stylesheet is. - List the manifests (`requirements.txt`, `package.json`, `pyproject.toml`). - Name any env-var fallback to leave alone, such as `API_URL || 'http://localhost:8000'`. - The orchestration and worktree state folder to skip.
Wersje
W bazie nie ma innych skilli o nazwie hk-code-quality-audit.
Migawka bazy z 13.11.2026 · wersja paczki e495d733bd07 (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
- 100unused
Sprawdza zadeklarowane zależności projektu, których żaden plik źródłowy nie importuje, i tylko raportuje wyniki. - 100run-ts-cleanup
Porządkuje bazę kodu TypeScript: usuwa martwy kod, nieużywane zależności, bezwartościowy kod z generatorów AI i słabe typy. - 84analyze-dead-code
Znajduje w repozytorium nieużywane eksporty, nieosiągalne moduły i martwe pliki, a następnie usuwa je z kodu. - 90finding-dead-code
Wyszukuje martwy i nieużywany kod, zbędne dodatki, osierocone fragmenty i nieużywane importy przy przeglądzie zmian, audycie nowych funkcji lub porządkowaniu. - 100hk-slop-audit
Przeprowadza audyt higieny kodu: szuka źle umieszczonych plików, martwego kodu, plików tymczasowych, problemów strukturalnych i zależnościowych oraz śmieci w git, i tworzy raport z priorytetami P0-P4. - 100knip-dead-code
Wykrywa martwy kod w JS i TS za pomocą Knip: nieużywane pliki, zależności, eksporty i typy, przy porządkowaniu kodu i kontroli zależności w CI.