Superstacks - Hub de Skills para Agentes de IA
KunanonJ/claude-skills-hub
6.485 skills portáveis para agentes de IA com foco em verificação e segurança
Ver no GitHubRepositório com 6.485 skills portáveis em formato SKILL.md para Claude Code, ChatGPT, Cursor, Gemini e outros agentes de IA. Fornece um stack enxuto focado em discovery, planejamento, TDD, verificação, review e pipelines de tickets com filosofia de 'ambiente sobre prompting'.
Instalação
npx convez-ai KunanonJ/claude-skills-hub1 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
- Potencializar agentes de IA (Claude, ChatGPT, Cursor) com skills reutilizáveis para desenvolvimento
- Implementar loops de verificação e TDD automatizados em fluxos de código
- Gerenciar pipelines de tickets e draft PRs com constraints de segurança
Veredito

Superstacks é um repositório robusto que codifica boas práticas de engenharia (verificação antes de escrita, constraints explícitas, safety overrides) em skills reutilizáveis. Ideal para equipes que querem elevar a qualidade de saída de agentes de IA sem depender apenas de prompts.
README
A tradução para o português deste README ainda está na fila.
Superstacks
Lean MIT skill stack for coding agents: discovery, plans, TDD, verification loops, review, draft PRs, and ticket pipelines.
Superstacks is a small plugin, not a giant catalog. It vendors a handful of upstream skills at pinned git SHAs, applies documented patches, and adds a generic ticket pipeline plus hard safety overrides. The GitHub repository name is KunanonJ/superstacks.
Philosophy
The engineer designs the environment. The model spends tokens on judgment.
- Environment over prompting. Put verification, constraints, and coordinators in the repo so the agent does not have to be reminded in prose.
- Verification first. Before writing code, name a deterministic check, run it red, then iterate until green. The agent should not be the human copying errors between tools.
- Constraints in code. When a mistake repeats, encode it as a type, lint rule, test, or directory convention. A rules line is last.
- Skills as process. A skill is a named loop with a script when the step is mechanical.
Themes drawn from the public Matt Pocock × Poteto conversation: YouTube. Paraphrase only; this README does not invent quotations.
Why this instead of raw upstream
- One folder you can list in Claude, Cursor, and Codex marketplaces.
- Descriptions stay short, trigger-first, and product-neutral in skill bodies.
- Safety overrides beat skill text: draft PRs,
Refs #N, evidence before "done", and atrust_levelladder (default 1: humans merge everything; a recorded owner decision may raise it, seeowner-decisions). - Vendored copies are hashed in
sources.lock.json. Patches live inpatches/.
Quick start
Pick one row. The GitHub repository is KunanonJ/superstacks.
| Surface | How |
|---|---|
| Claude Code | /plugin marketplace add KunanonJ/superstacks then /plugin install superstacks@superstacks |
| Cursor | Copy plugins/superstacks/ to ~/.cursor/plugins/local/superstacks/ until the public listing exists |
| Codex | codex plugin marketplace add KunanonJ/superstacks |
| claude.ai ZIP upload | python scripts/build_zips.py and upload a per-skill ZIP from dist/ (<skill>/SKILL.md) or dist/flat/ (SKILL.md at zip root) |
| ChatGPT ZIP upload | Same per-skill ZIPs, or dist/superstacks-plugin.zip for the OpenAI plugin portal |
npx skills add | DISABLE_TELEMETRY=1 npx skills add KunanonJ/superstacks -g -s '*' --copy -y |
| Plain copy | Copy plugins/superstacks/skills/<name>/ into your agent's skills directory |
A version bump on main is tagged and released automatically, with the ZIPs from dist/ attached (see CONTRIBUTING.md).
What this plugin runs, sends or fetches
Skills may tell the agent to run optional local helper scripts: git diff/git log, a user-supplied test command via verify-loop, or an existing Playwright install. The plugin itself makes no network calls, collects no telemetry, and does not read credentials. Optional installers such as npx skills add are third-party; pass DISABLE_TELEMETRY=1 if you use that path.
Pipeline
flowchart LR
ama --> how
how --> plans[writing-plans]
plans --> vloop[verify-loop]
vloop --> tdd[test-driven-development]
tdd --> verify[verification-before-completion]
verify --> review[code-review]
review --> optional[interrogate if one-way door]
optional --> pr[pr draft]
ticket-pipeline is the user-invoked router for that order; ama comes first and makes the assignment clear. verify-loop is the extra always-loaded skill: define the check, run it red, then green. safety-overrides always apply.
Off-pipeline (user-invoked): mistake-to-constraint, steer-miner, intake-triage, gardening, sample-review.
Skills
| Skill | Invocation | Source |
|---|---|---|
| verify-loop | model | this repo |
| owner-decisions | model | this repo |
| ama | model | this repo |
| writing-plans | model | obra/superpowers |
| test-driven-development | model | obra/superpowers |
| verification-before-completion | model | obra/superpowers |
| diagnosing-bugs | model | mattpocock/skills |
| code-review | model | mattpocock/skills |
| pr | model | mattpocock/skills |
| writing-for-agents | model | mattpocock/skills |
| grilling | model | mattpocock/skills |
| domain-modeling | model | mattpocock/skills |
| grill-with-docs | user | mattpocock/skills |
| how | user | pstack (cursor/plugins) |
| interrogate | user | pstack (cursor/plugins) |
| ticket-pipeline | user | this repo |
| workflow-profile | user | this repo |
| mistake-to-constraint | user | this repo |
| steer-miner | user | this repo |
| intake-triage | user | this repo |
| gardening | user | this repo |
| sample-review | user | this repo |
Always-on rule: plugins/superstacks/rules/safety-overrides.mdc.
Safety
- Draft pull requests. Never mark ready. Merge only as
trust_levelallows (default 1: humans merge everything). Authority to merge or mark ready comes from the owner's decision record (owner-decisions); with none, a human merges. Refs #NneverCloses/Fixes/Resolves.- Push the working branch only.
--force-with-leaseonly. - No production credentials, live infrastructure, runtime installs, or secret CI jobs.
- Plans go in the PR body. "Done" needs pasted command output.
- One-way doors (money, schema or data migrations, auth, data deletion, infra or prod config, licences) always need a human plus
interrogate.
See SECURITY.md.
Contributing
See CONTRIBUTING.md. Run python -m app.plugin_lint and python -m pytest before opening a pull request. Pull requests from the owner's account merge by themselves when the required check passes; pull requests from other people are safety-reviewed by an agent and then merged by the owner. See DECISIONS.md.
Credits
Vendored under MIT from Jesse Vincent, Matt Pocock, Lauren Tan / pstack, and a confirmed MIT excerpt of HumanLayer show-me inside pr. Full notices: plugins/superstacks/THIRD_PARTY.md.
License
MIT © 2026 Kunanon Jarat.