tdog Skills - Habilidades para Agentes de Coding
tylerdurrett/skills
Skills estruturadas para agentic coding com fluxo issue-tracker integrado
Ver no GitHubBiblioteca de skills para agentes de IA (Claude, Cursor, Codex) que implementa um fluxo completo de desenvolvimento: desde captura de specs, triage e decomposição até execução e deploy em branches de integração. Projetado para otimizar o ciclo de vida de features com rastreamento de issues no GitHub.
Instalação
npx convez-ai tylerdurrett/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
- Automatizar ciclo de vida de features (spec → triage → execução → ship)
- Decompor grandes requisitos em tasks managáveis para agentes
- Coordenar múltiplos agentes (Claude + Codex) em auditorias de código
- Manter specs sincronizadas com o tracker enquanto desenvolve
- Gerar recaps de stakeholder e status automático de progresso
Veredito

Conjunto bem estruturado e maduro de skills para quem quer automatizar o fluxo completo de desenvolvimento com agentes de IA. Requer setup inicial robusto (tracker, convenções de branch, docs), mas compensa em escala. Fortemente acoplado ao GitHub — não é para quem usa outros trackers.
README
A tradução para o português deste README ainda está na fila.
tdog skills
A library of agent skills for the way I build software with AI coding agents. Straight from my dev environment.
Install
npx skills@latest add tylerdurrett/skills
This uses the skills CLI by Vercel Labs. After running it, pick the skills you want and the agents you want them installed for (Claude Code, Cursor, Codex, etc.).
If the target repo doesn't already have a .claude/ directory, the CLI silently skips the Claude Code symlinks (it only writes the canonical .agents/skills/ copies). Create the directory first so the symlinks get made:
mkdir -p .claude && npx skills@latest add tylerdurrett/skills
First-time setup
Several of the engineering skills below (triage, execute, ship, audit, decompose, status, recap, to-spec, defer, check) assume a configured tracker, label vocabulary, integration-branch convention, and a docs/agents/ directory in your repo. Run setup-tdog-skills once after install to scaffold all of that:
/setup-tdog-skills
It walks you through two decisions (triage label vocabulary, single- vs multi-context domain docs), writes the canonical docs under docs/agents/, drops in the integration-branch ADR (docs/adr/0001-*.md by default, bumped to the next free slot if 0001 is already taken), and adds an ## Agent skills block to your CLAUDE.md / AGENTS.md. GitHub is the only tracker backend supported today; the workflow skills shell out to gh issue *. The non-tracker skills (tdd, diagnose, grill-me, code-simplifier, skill-creator, etc.) work standalone and don't need any setup.
The skills
Per-repo setup
| Skill | What it does |
|---|---|
| setup-tdog-skills | Scaffold the docs/agents/ directory, the integration-branch ADR, and the ## Agent skills memory block the workflow skills assume. Run once per repo. |
Spec lifecycle (issue-tracker workflow)
A four-tier flow — initiative → feature → slice → task — that publishes specs to GitHub Issues, triages them, decomposes them, executes them on per-spec integration branches, and ships them back up the chain.
| Skill | What it does |
|---|---|
| to-spec | Capture the current conversation as a sized spec (initiative / feature / slice / task) on the tracker. |
| triage | Verify size, lay down per-tier bookkeeping, apply the next state label. Also surfaces what's most actionable across the queue. |
| decompose | Break a tier-bearing spec into native sub-issues one tier smaller. |
| check | Fast read-only sanity check on a decomposition before execution burns a session on a flawed plan. |
| audit | Heavier multi-agent (Claude + Codex) version of /check, with writes back to the tracker on approval. |
| execute | Implement a size:task end-to-end on a branch off the parent's integration branch and open the PR. |
| ship | Tier-aware close-out: squash-merge tasks, promote slices/features upward, refuse initiatives (those close manually). |
| defer | Capture out-of-scope cleanup findings as cleanup-labeled issues so they don't pollute the current PR. |
| status | Read-only walk of the tracker that recommends one next step. |
| recap | Paste-ready stakeholder recap of recent activity (today / week / upcoming). |
Project docs
| Skill | What it does |
|---|---|
| grill-with-docs | Stress-test a plan against CONTEXT.md and docs/adr/, updating docs inline as decisions crystallise. |
| improve-codebase-architecture | Find deepening opportunities informed by CONTEXT.md and ADRs. |
| how-to-use | Static user manual: how the workflow flows, what skills exist, where to start. |
Coding workflow
| Skill | What it does |
|---|---|
| tdd | Red-green-refactor with integration-style tests. |
| diagnose | Disciplined diagnosis loop for hard bugs and perf regressions. |
| code-simplifier | Tighten recently-modified code without changing behaviour. |
| grill-me | Stress-test a plan or design via relentless interview. |
| zoom-out | Reorient when you're lost in a section of code. |
Skill authoring
| Skill | What it does |
|---|---|
| skill-creator | Guide for writing effective skills. |
| write-a-skill | Scaffold a new skill with progressive disclosure and bundled resources. |
Integrations & tools
| Skill | What it does |
|---|---|
| ai-sdk | Build features with the Vercel AI SDK (generateText, streamText, agents, RAG). |
| chrome-devtools | Browser automation, perf analysis, and debugging via Puppeteer. |
Credit
The install UX (npx skills@latest add owner/repo) is the vercel-labs/skills CLI. The "skills repo as a package" pattern is borrowed from mattpocock/skills, and several of the spec-lifecycle skills share lineage with his.
Licence
MIT.