Claude Skillz
SynapticSage/claude-skillz
Skills extensíveis para Claude Code com triggers por palavra-chave
Ver no GitHubColeção pessoal de skills para Claude Code que estendem funcionalidades com triggers baseados em palavras-chave. Inclui skills para análise adversarial, rastreamento de auditoria e integração com Codex via tmux, instaláveis localmente em ~/.claude/skills/.
Instalação
npx convez-ai SynapticSage/claude-skillz1 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
- Stress-testing e análise adversarial de decisões críticas
- Geração de relatórios de auditoria e rastreamento de mudanças
- Integração com OpenAI Codex como colega de pair programming
Veredito

Repositório técnico bem estruturado com skills prontas para produção, mas voltado a usuários avançados que entendem tmux, shell e architecture de Claude Code. Documentação clara e convenções bem estabelecidas facilitam extensão própria.
README
A tradução para o português deste README ainda está na fila.
claude-skillz
Personal collection of Claude Code skills,
installed at ~/.claude/skills/. Each subdirectory is a self-contained
skill that Claude Code loads automatically at startup.
What's here
| Skill | Triggers when the user says… | What it does |
|---|---|---|
adversarial | "stress-test", "find holes", "pressure-test", "devil's advocate", "red team", "challenge this", "check my reasoning" | Runs parallel adversarial and counter-adversarial agents across multiple rounds, with explicit convergence detection. Yields a synthesis that distinguishes well-argued conclusions from robust ones. Built for high-stakes decisions where confirmation bias is dangerous. |
audit-trail | "audit trail", "recap", "trace", "work log", "show your work", "what changed", "what did you do" | Produces a bounded, evidence-first report of recent work: user requests, actions taken, files changed, decisions, verification, risks, next steps. Supports a positional budget arg (80, 500w, 3000c) and flags for included sections, plus opt-in ASCII/HTML change-flow diagrams. |
codex-pair | "pair with codex", "ask codex", "consult codex", "second opinion from codex", "codex review", "codex challenge" | Spawns and manages OpenAI Codex CLI as a persistent teammate in a sibling tmux pane. Sends prompts, captures responses, preserves Codex session memory across turns. Uses the ChatGPT subscription (codex login), not the OpenAI API. Optional tmux-bridge-mcp transport enables bidirectional Codex↔CC calls — see codex-pair/install.sh. |
How Claude Code discovers skills
Claude Code scans ~/.claude/skills/ at startup and reads each
subdirectory's SKILL.md frontmatter (name, description,
allowed-tools). The description is what Claude uses to decide when a
skill applies — it should list concrete trigger phrases the user might
actually say.
When a skill fires, Claude Code loads the full SKILL.md body into the
conversation as operating instructions.
No build step, no registration file, no plugin manifest. Drop a directory in; it's a skill. Remove it; it's gone.
Skills that need extra setup
Most skills here are pure prompt/markdown and work as soon as they land on
disk. codex-pair is the exception — it wraps an external CLI (codex)
and optionally a local MCP server (tmux-bridge-mcp), so it ships an
install.sh that handles building and registering those runtime
dependencies idempotently. See its own README/SKILL.md for details.
Layout
~/.claude/skills/
├── README.md ← this file
├── .gitignore ← skill-local scratch excluded from version control
├── adversarial/SKILL.md
├── audit-trail/SKILL.md
└── codex-pair/
├── SKILL.md ← invocation instructions
├── install.sh ← idempotent + self-healing setup
└── TODO.md ← known debt / open issues
Conventions these skills follow
- Trigger phrases live in the
descriptionfrontmatter, not the body. Claude matches on the description when deciding whether a skill applies. - Body is operating instructions for Claude, not user-facing docs. Use the imperative mood: "Read the file", "Do X", "Stop if Y."
- Back up before destructive changes. When a skill edits shared state
(
settings.json, long-lived tmux panes, config files), it should detect existing state, refuse to blindly overwrite malformed content, and leave an escape hatch (a.bak, a git commit, an--uninstallflag). - Prefer idempotent + self-healing. A skill should be safe to run twice; a second run on drifted state should restore it, not error.
- Be specific about triggers. Vague descriptions ("general helper")
make the skill fire on the wrong prompts. Concrete phrases in the
descriptionfrontmatter let Claude distinguish it from adjacent skills.
License & provenance
Personal repo — no warranty, no stability guarantee. If something looks
useful, copy it into your own ~/.claude/skills/ and adapt.