convez.ai
SkillAutomaçãoAvançado
dzianisv

Skills Reutilizáveis para Agentes Claude

dzianisv/skills

Skills agnósticas para agentes Claude, OpenCode e compatíveis

Ver no GitHub
InicialPluginScore · de 0 a 100

Biblioteca de skills (competências) reutilizáveis para agentes como Claude Code, OpenCode e Codex, seguindo a especificação Agent Skill. Cada skill é um workflow encapsulado que o agente carrega automaticamente conforme o contexto, cobrindo desde automação de browser até integração com serviços como Google Workspace e Vercel.

Instalação

$npx convez-ai dzianisv/skills

1 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

Estrelas
4
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 dia
Versão
—
Issues abertas
2
Contribuidores
3

Casos de uso

  • Automação de navegador com controle de Chrome persistente
  • Fluxos end-to-end de desenvolvimento (design, implementação, review, merge)
  • Integração com serviços de deploy, DNS dinâmico e Google Workspace
  • Verificação de saúde de serviços e orquestração de componentes OpenCode

Veredito

Projeto maduro e bem estruturado para estender capacidades de agentes IA. Atende especialmente quem precisa de automação complexa e integração com ferramentas externas, mas requer familiaridade com especificações de agent skills e CLI.

README

A tradução para o português deste README ainda está na fila.

skills

Reusable agent skills for Claude Code, OpenCode, Codex, and any other Anthropic/OpenAI-compatible agent runtime that consumes the Agent Skill spec.

Each skill is a single directory under skills/ containing a SKILL.md (and optional scripts/references). Skills are platform-agnostic — they describe a workflow or technique with a name + description frontmatter that tells the agent when to load it.

Quickstart

Install all skills at once:

npx -y skills add dzianisv/skills -g

Or install a single skill by name:

npx -y skills add dzianisv/skills -s <skill-name> -g
# example:
npx -y skills add dzianisv/skills -s my-browser -g

The -s flag preselects the skill; without it the CLI prompts you to pick one. -g installs globally for your runtime; drop it (or use -p) to install into the current project.

After install, the agent autoloads a skill when its description/triggers match your request, or you can invoke it explicitly (e.g. /my-browser, /own).

Skills in this repo

SkillWhat it doesInstall
my-browserControl your real running Chrome (live sessions, cookies, auth) via a persistent Puppeteer gateway using autoConnect mode. No --remote-debugging-port needed.npx -y skills add dzianisv/skills -s my-browser -g
chrome-useDrive your real running Chrome with agent-browser-style commands (open, snapshot -i, click @e1, fill, screenshot). Zero-dependency TypeScript speaking CDP directly over the built-in WebSocket; connects via autoConnect like my-browser.npx -y skills add dzianisv/skills -s chrome-use -g
ownTake full ownership of a task end-to-end: issue → design → plan → implement → review → real-feature test → PR → CI → final review → merge ask. Forbids back-delegation to the user when a tool or credential store can resolve the blocker.npx -y skills add dzianisv/skills -s own -g
duckdns-domainRegister a free DuckDNS subdomain and point it at a server IP via Chrome DevTools browser automation. Useful when you need a public DNS name for a dev server and no managed DNS is available.npx -y skills add dzianisv/skills -s duckdns-domain -g
readiness-checkVerify all OpenCode plugin services (Whisper, TTS, Supabase, Telegram, etc.) are healthy and ready.npx -y skills add dzianisv/skills -s readiness-check -g
vercel-lfsDeploy a site with Git LFS assets (videos, images) to Vercel without serving 15-byte pointer files. Covers the prebuilt-deploy workflow and required GitHub secrets.npx -y skills add dzianisv/skills -s vercel-lfs -g
skills-shPublish, register, and troubleshoot skills on skills.sh, the public agent-skills directory.npx -y skills add dzianisv/skills -s skills-sh -g
opencode-apiControl OpenCode through the REST API exposed by opencode serve. Sessions, prompts, SSE events, permissions, and automation.npx -y skills add dzianisv/skills -s opencode-api -g
google-workspace-cliInteract with all Google Workspace APIs (Drive, Gmail, Calendar, Sheets, Docs, Chat, etc.) via the gws CLI.npx -y skills add dzianisv/skills -s google-workspace-cli -g
claude-sessionsFind a past Claude Code session by project or topic and summarize what was done in it — "which session did I work on X in?". Cheaply ranks local transcripts, then fans the heavy reads out to subagents so multi-MB sessions never flood your context.npx -y skills add dzianisv/skills -s claude-sessions -g
solo-founderOperate as an autonomous solo-founder / CEO meta-agent: decide what matters, spawn doer subagents across product / growth / distribution, evaluate + improve them, ship and verify. Bundles a cross-session run registry (PreToolUse hook) + an LLM-judge evaluator so its runs can be scored later.npx -y skills add dzianisv/skills -s solo-founder -g
solo-founder-meta-agentEvaluate and improve the solo-founder skill from its REAL recent runs (the ~/.local/run/solo-founder/ registry + OpenCode sessions): score each session, diagnose the dominant failure, harvest it into a frozen eval case, and run the DGM-H improve-loop so a meta agent rewrites SKILL.md.npx -y skills add dzianisv/skills -s solo-founder-meta-agent -g
take-ownership-meta-agentEvaluate and improve the take-ownership skill from its REAL recent runs (~/.local/run/take-ownership/ registry). Scores six dimensions (r1_defined, no_fake_done, phase_discipline, real_testing, state_persisted, blocker_resolved), diagnoses the dominant failure, harvests frozen eval cases, and runs the DGM-H improve-loop. Bundles 5 train cases, 2 holdout cases, a driver (driver-to.py), and a PreToolUse tracking hook.npx -y skills add dzianisv/skills -s take-ownership-meta-agent -g
git-worktreeCreate a git worktree for isolated branch work by symlinking node_modules and .env files from the main repo — no reinstall. Resolves the main repo's directory name dynamically (no hardcoded path).npx -y skills add dzianisv/skills -s git-worktree -g

How claude-sessions works

Claude Code stores every session as a JSONL transcript under ~/.claude/projects/<dashified-cwd>/<sessionId>.jsonl (the directory name is the project's working directory with / and . replaced by -). These files grow to tens of MB, so reading them directly blows up an agent's context.

This skill avoids that in three steps:

  1. Rank cheaply. scripts/rank_sessions.py makes a single streaming pass over each transcript, scoring it by keyword hits in its title / last prompt / first message plus a whole-file hit count, and prints the top candidates as JSON. It handles sibling project dirs (…-OpenClawBot, …-OpenClawBot-2) and supports --project, --days, and --limit.
  2. Deep-read with subagents. The agent dispatches one subagent per top candidate, in parallel — each greps within its assigned file and reads only the surrounding lines, then reports whether it matches and a short summary. The large transcripts stay out of the main context.
  3. Present the best match with its summary, session_id, path, and date (resume it with claude --resume <session_id>).

Try the ranker directly:

python3 ~/.claude/skills/claude-sessions/scripts/rank_sessions.py \
  --project ~/workspace/OpenClawBot "merge gate" enforce testing

Or just ask your agent: "Find the session where I worked on PR merge-gate enforcement testing in ~/workspace/OpenClawBot and summarize it."

Install

Pick whichever location your runtime expects:

# Claude Code (global skills)
git clone https://github.com/dzianisv/skills.git /tmp/skills
cp -r /tmp/skills/skills/* ~/.claude/skills/

# OpenCode
cp -r /tmp/skills/skills/* ~/.config/opencode/skills/

# Codex / generic
cp -r /tmp/skills/skills/* ~/.agents/skills/

Or symlink individual skills:

ln -s "$(pwd)/skills/own" ~/.claude/skills/own

After install, the agent will autoload a skill when its description/triggers match the user's request, or you can invoke it explicitly (/own, /duckdns, etc., depending on your runtime's slash-command conventions).

Spec

These skills follow the agentskills.io spec. Frontmatter:

---
name: <kebab-case-name>
description: Use when <triggering conditions and symptoms>
---

Contributing

PRs welcome. New skills should:

  • Have a clear description field that starts with "Use when…" and describes triggering conditions, not the workflow itself.
  • Be platform-neutral where possible. If a skill is tied to a specific runtime (Claude Code vs. OpenCode), say so in the body.
  • Avoid hard-coded personal paths, account names, IPs, or secrets.

License

MIT — see LICENSE.

Ver o original no GitHub

Bom para

Relacionados