24 skills · 3 pacotes

Claude Code skills,
curadas.

Skills do meu setup do Claude Code, em três pacotes — build (o método: do problema ao commit), ship (a entrega: Jira, PR, homolog e produção) e ferramentas avulsas. Instale o pacote, cole os comandos no seu Claude Code e pronto: funciona em qualquer sistema e as atualizações chegam automáticas.

Instale um pacote — já vem com todas as skills dentro

Como as skills se puxam

Cada skill declara suas relações no próprio frontmatter — este mapa é desenhado a partir delas. Embaixo ficam as fundações; quem depende sobe.

  • Build
  • Ship
  • Toolbox
First truly understands the problem and the project, then writes the card as PM/PO and creates it on the right board with destination, owner and images. With other skills in the argument it understands and combines them in one pass. Refuses without Jira./cardDesign doctrine, UX and UI: #1 reference, wow, anyone uses it unaided, every screen by clicks and URL; raises a target, behaviour unchanged, `audit` = report only; never commits nor branches./frontHomolog live with what shipped: /pull-request guaranteed → PR approved (reviewed and fixed if you are not the author) and merged → environment configured from the infra map → checked live → Jira card → return. Never touches production./homologThe infra map of THIS project in `.claude/ship/infra.md`: accounts, accesses, where each secret lives (never a value), homolog and prod environments. `audit` reports only./infraThe rigorous engineering protocol: 11 steps (plus Step 0) from problem to committed code, one artifact per step, /solve reinvoked every step./methodThe engineering doctrine: SOLID (all five), DRY, KISS, YAGNI, Law of Demeter, Motores (one owner per capability) and continuous refactoring (everything the work touches goes up). With a target (folder, file, diff, commit) it raises it without changing behaviour; `audit` = report only. Never commits, never creates a branch./principles/homolog guaranteed → homolog promoted (PR reviewed and fixed, or merge and push) → configured, checked live, Jira card. Single owner of production./prodUse when user invokes /proto to recreate a screen in 3 versions, each on a temporary parallel route in the same app ({rota-original}-v1/-v2/-v3), respecting the app design system, mobile and desktop, at /solve quality — as the screen SHOULD be, not as it is. Ends with the 3 URLs and a recommendation so the user picks one; the chosen version is implemented later by /method./protoPublishes the work: /work complete → target merged in → push → PR where the setup says (or push only) → title, description, comment → Jira card → return. Never merges./pull-requestResolve the requested task at world-class level, benchmarking against the leading big pop tech apps in the relevant domain as the quality baseline, aiming to make us the #1 reference in the market. Invokes /principles (the engineering doctrine) and, with a visual surface, /front (the design doctrine) via the Skill tool./solveUse when user invokes /sync with a branch expression like "main > dev", "dev = main" or "A = B > C gh = local" — a notation for synchronizing git branches: `=` converges to the same commit (nobody loses work), `>` flows one way (right never flows back), `gh = local` means local and GitHub identical. Conflicts resolved by understanding both sides; never --force, reset --hard, rebase or branch deletion./syncThe ground (setup, card, branch, status), then the commit with the front screens proven on the card. Never pushes./workReproduces the report on the front as the user, judged ≥ 90; the dev sees the bug and the fix BEFORE the commit./reproUse when user invokes /ask to call the AskUserQuestion tool inline during the current task./askUse when user invokes /blind <texto> — the text goes verbatim to a fresh Claude session that sees nothing (no conversation, CLAUDE.md, memory, files, internet or tools) and the answer comes back whole. Removes context bias, not model bias./blindUse when user invokes /brain to understand a problem or decide between paths — thinking, not building. Assumes the request is WRONG and restates it; seven phases, blind judge on one-way doors, ONE recommendation as an ADR in docs/decisions/./brainUse when user invokes /chat — activates read-only guru mode: deep analysis and smart answers with no file or system modification. The mode ends when the user types `chat out`; nothing else ends it./chatUse when user invokes /claude-shortcuts to set up, verify or replicate the user's Claude Code shortcuts: terminal `claude` / `claudew` (ultracode) and IDE terminal profiles "Claude" (Ctrl+Q, --effort max, tunable) and "Claude Ultra" (Ctrl+Shift+U, ultracode) for VS Code + Antigravity IDE, across all profiles./claude-shortcutsUse when user invokes /create-test [tema] — supre a pirâmide de teste do tema tratado no prompt ou na conversa: unidade, integração e E2E no Playwright, no padrão do projeto e rodados até verde./create-testUse when user invokes /ctt to capture a request into the Claude Code task list (TaskCreate) instead of doing it now — an inbox for ideas that surface mid-conversation without derailing the current work./cttUse when the user types /make-dev — makes `make dev` the one command that runs this project: fresh clone, port conflicts, "another dev server is already running", Makefile updates, hot reload or auto-migrations that should just work, and Actions on the local self-hosted runner./make-devUse when user invokes /make-skill <what the skill must do | path to a SKILL.md> — writes (or rewrites) one SKILL.md under the rule: simple, clean, minimal, readable by a human and by the model, efficient, written as an order (never a how-to), under 3333 chars; places it where this repo keeps skills and proves it./make-skillUse when user invokes /save in a git repo to commit current work, with optional override message in Conventional Commits format/saveUse when the user runs /video-teams <link> to download a Microsoft Teams / SharePoint / OneDrive "Stream" recording (stream.aspx, *.svc.ms videomanifest) that plays in the logged-in Playwright browser but has download blocked — AES/SEA-encrypted DASH that yt-dlp/ffmpeg cannot fetch. Pulls video + transcript (srt/vtt) to ~/Downloads./video-teams

Toque numa skill para acender as relações dela e prendê-la aqui. Arraste o mapa para os lados.