AI Tooling Free - Skills e Agents Multiplataforma
nortonx/ai-tooling-free
Skills e agents reutilizáveis para IA em múltiplas plataformas
Ver no GitHubConjunto de 16 skills e 10 agents em PowerShell para Claude Code, GitHub Copilot CLI, Gemini CLI e padrão Agent Skills. Cobre code review, testes, refatoração, documentação e otimização sem dependências pesadas. Instalação por symlinks, funciona em macOS, Linux, WSL2 e Windows sem direitos admin.
Instalação
npx convez-ai nortonx/ai-tooling-free1 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
- Automação de revisão de código com múltiplas perspectivas
- Geração automática de testes unitários e mensagens de commit
- Auditoria de regras ESLint, Prettier e Biome
- Detecção de violações DRY e guias de refatoração
- Scaffold de ADRs, especificações e testes de cobertura
Veredito

Projeto maduro com abordagem pragmática: symlinks ao invés de cópias, sem mágica npm/postinstall, compatível com o padrão aberto Agent Skills. Ideal para equipes que usam múltiplos clientes IA e querem um único repositório de origem.
README
A tradução para o português deste README ainda está na fila.
ai-tooling-free
Curated, cross-platform AI coding skills and agents for Claude Code, GitHub Copilot CLI, Gemini CLI / Antigravity, and any Agent Skills standard adopter (OpenAI Codex CLI, Cursor, Amp, OpenCode, …) — one source of truth, symlinked into all of them.
16 skills, 10 agents, a dependency-light setup script. No npm install, no postinstall magic, no settings rewriting.
Requirements
- One or more of: Claude Code, GitHub Copilot CLI, Gemini CLI / Antigravity, or any Agent Skills adopter (OpenAI Codex CLI, Cursor, …)
git(recommended — several skills are git-centric)- Keep the cloned repo where it is: skills are symlinked, not copied. If you move the clone, re-run setup.
Install
macOS / Linux / WSL2
git clone https://github.com/nortonx/ai-tooling-free
cd ai-tooling-free
./setup.sh
Windows
git clone https://github.com/nortonx/ai-tooling-free
cd ai-tooling-free
setup.cmd
Windows uses directory junctions (mklink /J) — no admin rights or Developer Mode needed. setup.cmd works with both Windows PowerShell 5.1 and pwsh 7+.
What setup does (and doesn't)
Does:
- Symlinks each skill into
~/.claude/skills(Claude Code) and~/.agents/skills— the latter is the canonical user dir of the Agent Skills open standard, so one link target covers Copilot CLI, Codex CLI, Cursor, Gemini CLI / Antigravity, and every other adopter. (Skills are not also linked into~/.copilot/skills— Copilot reads~/.agents/skillstoo, and since it reads both roots without de-duplicating it would list every skill twice; nor into~/.gemini/skills— Gemini treats~/.agents/skillsas a same-tier alias that takes precedence. Setup prunes any~/.copilot/skillsand~/.gemini/skillslinks left by older installs.) - Links each agent into
~/.claude/agents(Claude Code only; copied on Windows) - Backs up anything already at a destination to
<name>.bakbefore linking - Is idempotent — re-run it any time
Doesn't:
- Touch your
settings.json, model, theme, permissions, or plugins - Deploy any global
CLAUDE.md/ instruction files - Delete anything (backups only)
- Require
jq, Node, or anything beyond a shell
Skills
| Skill | What it does |
|---|---|
fanout-review | Multi-perspective code review: 6 parallel reviewers + deterministic merge gate, output ready to paste into a PR thread |
check-dx | Audit your eslint / prettier / biome rules and rank keep / tune / drop via consensus across 5 independent evaluation passes |
framework-upgrade-guide | Stepwise major-by-major upgrade guide for Angular / React / Vue projects, with lockstep version matrices and per-hop verification |
check-dry | Detect DRY violations and recommend refactoring strategies |
check-tests | Check test coverage for the changes on the current branch |
create-commit-message | Ready-to-paste conventional commit message from your current changes |
create-unit-tests | Write unit tests for branch changes following best practices |
generate-adr | Scaffold an Architecture Decision Record (Nygard format) with auto-numbering and supersedes links |
generate-spec | Scaffold a feature spec using a reusable 9-section template |
learn | Explain a concept in depth — ELI5, technical, analogies, pitfalls |
optimize | Find performance bottlenecks and recommend optimizations |
plan-or-execute | Recommend plan mode vs direct execution for a task |
pr-description | Generate a PR description from commits and diff |
ship-it | Branch (if needed), commit, and push in one reviewed step |
smart-fix | Route an issue description to the right flow: bug, performance, security, or feature |
update-claude-md | Update a project's CLAUDE.md with conventions learned from recent changes |
Skills work in Claude Code, Copilot CLI, Gemini CLI, and Agent Skills adopters like Codex CLI and Cursor (same SKILL.md format).
Cross-tool caveats
$ARGUMENTSonly substitutes in Claude Code — every other tool's skill body sees the literal text. Pass arguments inline in your prompt; each skill documents this in its## Argumentssection.- Skill bodies use Claude Code tool names (Read, Edit, Grep, Agent, …). Other tools map them to their own equivalents loosely — most skills work fine, but agent-dispatch steps are Claude-only.
- Codex CLI: list skills with
/skillsor$-mention; disable individual skills via~/.codex/config.toml. This repo does not create~/.codex/AGENTS.mdor any global instructions. - Copilot and Cursor scan both
~/.claude/skillsand~/.agents/skillswithout de-duplicating, so they list each skill twice. This is inherent, not a setup bug: Claude Code reads only~/.claude/skillsand Codex reads only~/.agents/skills, so both roots must stay populated, and Copilot/Cursor read both. Removing~/.copilot/skillsbrings Copilot down from three entries to two; reaching exactly one would mean dropping Claude Code or waiting on Copilot to de-dup (github/copilot-cli#2161). The duplicate links resolve to the same target, so it's cosmetic. - Skill discovery in Codex/Cursor hasn't been smoke-tested by this setup — the links follow the documented Agent Skills locations; verify with
/skills(Codex) or the Agent skill list (Cursor) after install.
Agents (Claude Code only)
communication-excellence-coach · debugger · doc-writer · mermaid-diagram-specialist · performance-optimizer · refactoring-expert · security-auditor · system-architect · test-automator · ui-ux-designer
Agents are subagents — specialist personas Claude Code delegates to in their own context window. Unlike skills, they have no slash command, and only Claude Code supports them (Copilot CLI, Gemini, and Codex have no subagent mechanism, so setup doesn't link agents into them).
Two ways to use one:
-
Automatic — Claude reads each agent's
descriptionand delegates when the task fits (several are marked "Use PROACTIVELY"). -
Explicit — name the agent in your prompt:
> Use the debugger subagent to find why this test fails
Manage installed agents with the /agents command.
Uninstall
Setup only creates links (plus .bak backups) — nothing of yours is deleted. To remove:
# remove the skill symlinks this repo created
for d in ~/.claude/skills ~/.copilot/skills ~/.gemini/skills ~/.agents/skills; do
find "$d" -maxdepth 1 -type l -lname "*/ai-tooling-free/*" -delete
done
# remove the agent links
find ~/.claude/agents -maxdepth 1 -type l -lname "*/ai-tooling-free/*" -delete
On Windows, delete the junctions under %USERPROFILE%\.claude\skills (and .copilot, .gemini, .agents) — rmdir <name> removes a junction without touching this repo. Restore any .bak files you want back.