Flywheel Skills
We-The-Flywheel/skills
10+ skills prontos para Claude Code: escrita, raciocínio, QA, decisões e workflow
Ver no GitHubColeção de skills reutilizáveis para Claude Code, distribuída via marketplace de plugins. Inclui ferramentas de escrita (humanizador, verificação de idiota), raciocínio multi-modelo, QA visual, planejamento de pré-mortem, automação de workflow e gates de qualidade — instale tudo com dois comandos.
Instalação
npx convez-ai We-The-Flywheel/skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Remover artefatos de escrita gerada por IA antes de publicar
- Executar deliberação multi-modelo para decisões arquiteturais complexas
- Testar responsividade visual de sites em múltiplos dispositivos
- Validar planos antes de execução via análise de falhas potenciais
- Automação de checklist de qualidade e release gates baseados em evidência
Veredito

Suite bem pensada e prática de skills que endereça gargalos reais (escrita de IA, decisões em grupo, QA visual, workflow). Instalação sem atrito via marketplace. Excelente para equipes que já trabalham com Claude Code.
README
A tradução para o português deste README ainda está na fila.
Flywheel Skills
Reusable Claude Code skills maintained by The Flywheel.
Distributed as a Claude Code plugin marketplace so you can install everything with two commands — no manual file copying.
Install (recommended — plugin marketplace)
In any Claude Code session:
/plugin marketplace add We-The-Flywheel/skills
/plugin install wtf@flywheel
That's it. The skills are now available, namespaced under wtf::
wtf:humanizerwtf:multi-llm-deliberationwtf:visual-qawtf:premortemwtf:endwtf:idiocy-checkwtf:release-gatewtf:content-gatewtf:moodboardwtf:website-build
To update later: /plugin marketplace update flywheel.
Install (alternative — clone + script)
If you'd rather not use the marketplace (older Claude Code, or you just prefer cloning):
git clone https://github.com/We-The-Flywheel/skills.git
cd skills
./scripts/install.sh
install.sh copies each skill into ~/.claude/skills/ with a wtf- prefix
(wtf-humanizer, wtf-multi-llm-deliberation, wtf-visual-qa, wtf-premortem, wtf-end, wtf-idiocy-check, wtf-release-gate, wtf-content-gate, wtf-moodboard, wtf-website-build) so they never clash
with same-named skills you may already have. Re-running it skips anything already
installed.
The skills
Grouped by category. (Claude Code discovers plugin skills one level deep, so they live
flat under plugins/wtf/skills/ — the categories below are organizational, not directories.)
| Category | Skill | What it does | Extra setup |
|---|---|---|---|
| Writing | humanizer | Strips the tells of AI-generated writing (em-dash overuse, rule-of-three, inflated symbolism, vague attributions, …). Optionally matches a per-project VOICE.md. | None |
| Reasoning | multi-llm-deliberation | Runs a 3-stage deliberation (diverge → rank → synthesize) across multiple models via OpenRouter for consensus answers on architecture, code review, or hard questions. Includes a Content Truth-Check Mode that fact-checks draft articles: atomic claim extraction → cross-model verdicts (disagreement = hallucination flag) → web verification of volatile/disputed claims → surgical fixes. | OPENROUTER_API_KEY in your environment or ~/.env.shared |
| QA | visual-qa | Captures full-page screenshots of a site at desktop/tablet/mobile widths and renders them into one tabbed HTML gallery for visual review. | Node.js; runs npm install (Playwright) on first use via setup.sh |
| Decisions | premortem | Stress-tests a plan before you commit: imagines it's failed months from now, spawns one investigator per failure mode in parallel, then synthesizes the most likely / most dangerous failure, the biggest hidden assumption, a revised plan, and a pre-commit checklist. | None |
| Workflow | end | Wraps up a coding session: shuts down local dev servers, removes temp/backup files, commits and pushes outstanding work, and refreshes project docs (PROJECT_MAP.md + CLAUDE.md). Safe-by-default — confirms before anything destructive. | None |
| Writing | idiocy-check | Fast, ruthless pre-submission review of any document, grant, caption, email, or deliverable. Returns 5–8 items that would embarrass you, get you rejected, or make you look sloppy — not a comprehensive edit. Contributed by Eric Cross. | None |
| QA | release-gate | Evidence-based ship gate for larger implementations: deterministic gates (secrets scan, build, lint, tests, coverage delta), rubric'd pass/fail dimension checks with adversarial verification of every finding, and runtime evidence (run the app, observe). Emits a PASS/FAIL report card. Read-only — never edits. Supports a per-project VERIFY_RUBRIC.md. | None |
| QA | content-gate | Nine-step pre-publish gate for web content (blog posts, landing pages, SEO articles): draft quality + voice profile, fact-check, de-AI pass, hero/OG image, full OG/Twitter meta, FAQ + FAQPage JSON-LD, schema + E-E-A-T signals, analytics coverage, and AI-citation/zero-click readiness (self-contained answer above the fold — ~60% of searches end without a click). Emits a per-step PASS/FAIL report card; any FAIL blocks publish. Pairs with wtf:humanizer (Step 3) and wtf:multi-llm-deliberation Truth-Check Mode (Step 2) when installed. | None |
| Design | moodboard | Turns visual references (and anti-references) into locked design decisions — background, type category, accent approach, nav scale, interaction patterns, anti-patterns — each traceable to a finding. Includes an in-context type explorer (renders the real wordmark in 8–12 fonts on the actual brand background, one scroll). Produces documented decisions, not pages. | None |
| Design | website-build | Reference-first site build on a four-pillars framework (Audience → Structure → Copy → Design): clarity-paragraph gate, audience filter, four-visitor-jobs structure with an accountability map for every cut, voice spec, and a Design pillar handed to wtf:moodboard. Produces a populated brief; blanks are the to-do list. Pairs with wtf:humanizer and wtf:content-gate when installed. | None |
When to use what — the lifecycle
The skills are designed to slot into the phases of a normal product loop. You don't need all of them on every change — match the phase you're in:
PLAN ──────────► BUILD ──────────► CHECK ──────────► SHIP ──────────► REFLECT
premortem (Claude Code code review* release-gate retro habits
multi-llm- plan mode, visual-qa content-gate (what failed →
deliberation tests as humanizer end new rubric line
you go) idiocy-check or skill)
- PLAN — before committing to an approach:
wtf:premortemstress-tests the plan (how does this fail?);wtf:multi-llm-deliberationsettles architecture calls when there are >2 defensible approaches. - BUILD — mostly Claude Code itself: plan mode, tests alongside code.
- CHECK — two distinct steps, in order:
- Code review (judgment — Claude Code's built-in
/code-review*): findings, opinions, suggestions. "Is this good code?" wtf:release-gate(evidence — this pack): binary gates with proof. "Does this meet the bar to ship?" Review advises; the gate gatekeeps. Pluswtf:visual-qafor anything with a UI, andwtf:humanizer/wtf:idiocy-checkfor prose and deliverables.
- Code review (judgment — Claude Code's built-in
- SHIP —
wtf:release-gatemust be green first; for anything going to a public URL (blog posts, landing pages, SEO articles),wtf:content-gatemust also be green — it's the content counterpart to the release gate (meta tags, FAQ + structured data, E-E-A-T, analytics, zero-click/AI-citation readiness). Then commit/push/deploy and close the session withwtf:end. - REFLECT — when the gate or review caught something late, encode it: add a line to
your
VERIFY_RUBRIC.mdso the gate catches it automatically next time.
* /code-review is built into Claude Code, not part of this pack — listed for the
sequence's sake.
Contributing
New skills welcome — see CONTRIBUTING.md. Every change is gated by an automated leak-check that blocks internal hostnames, private paths, and credentials from ever landing in this public repo.
License
MIT (see LICENSE). Bundled third-party work is attributed in NOTICE.