Claude Tools - Agentes e Skills para Desenvolvimento
Sebstrdigital/claude-tools
Agentes e skills modulares para automação de tarefas técnicas com Claude
Ver no GitHubColeção de agentes agentic customizados e skills reutilizáveis para Claude, cobrindo auditorias, reviews de código, design de harness, orquestração de trabalho e produtividade. Inclui agentes especializados (scout, builder, heavy, skeptic, grunt) e skills modulares em engenharia, produtividade e utilitários, instaláveis via script automatizado.
Instalação
npx convez-ai Sebstrdigital/claude-tools1 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
- Auditorias de segurança, integridade de dados e arquitetura agentic
- Reviews profundos de código com análise de padrões e anti-padrões
- Design e implementação de harness para sistemas autônomos escaláveis
- Orquestração de trabalho técnico delegando entre agentes especializados
- Investigação de codebase, debugging e refactoring com reconhecimento
Veredito

Repositório maduro e bem-estruturado de ferramentas enterprise para quem constrói sistemas agentic complexos. Cobertura robusta em auditorias, reviews e padrões, mas requer familiaridade com arquitetura de agentes e CLI; não é para inicialização rápida.
README
A tradução para o português deste README ainda está na fila.
Agentic Tools
Custom agentic AI agents, Claude Code skills, and reference docs by @Sebstrdigital.
What's Included
Agents
Active agents live in agents/ and are installed by ./install.sh.
| Agent | Description |
|---|---|
scout | Read-only reconnaissance worker for codebase investigation, execution tracing, external repo/doc surveys, and evidence-backed findings. |
builder | Standard implementation worker for scoped stories, fixes, tests, and refactors with clear validation commands. |
heavy | Senior implementation worker for hard debugging, ambiguous root-cause hunts, cross-cutting refactors, and security-sensitive work. |
skeptic | Adversarial read-only reviewer that tries to refute diffs, claims, and validations before work is accepted. |
grunt | Mechanical-work worker for low-judgment tasks such as renames, doc syncs, formatting sweeps, bulk searches, and known-file installs. |
Retired repo-managed agents are removed from installed Claude Code agents by install.sh when they contain this repo's source_id.
Skills
Skills live in category-first folders under skills/ and are installed by ./install.sh to ~/.claude/skills/<skill-name>/. This follows the Matt Pocock-style layout:
skills/
├── engineering/
├── productivity/
└── misc/
Source paths are:
skills/engineering/<skill-name>/skills/productivity/<skill-name>/skills/misc/<skill-name>/
Each skill has a required SKILL.md. Supporting files live beside it.
Engineering
| Skill | Description |
|---|---|
agent-audit | Quick pass/fail pre-deployment audit for agentic systems. |
agent-review | Deep agentic architecture review with pattern scoring, anti-pattern scanning, and recommendations. |
code-review | General code quality review across structure, naming, complexity, error handling, and type safety/testing. |
data-integrity-audit | Audit AI-generated query results from user question through displayed answer. |
deep-review | Four-pass code review chain covering framework correctness, resource safety, security, and adversarial absence detection. |
harness-interview | Guided interview capturing domain, systems, platform, autonomy, scale, and harness-type signals. |
harness-architect | Selects harness type, designs agent decomposition, orchestration, and tool map. |
harness-spec | Generates an implementation-ready specification for a harness. |
harness-build | Generates runnable scaffolding adapted to harness type and platform. |
harness-review | Deep review of harness implementations against harness patterns and anti-patterns. |
infra-audit | Interview-based infrastructure assessment with gap analysis and migration roadmap. |
patterns | Diagnose problems and analyze architectures against agentic design patterns. |
project-audit | Full audit orchestrator that runs applicable audits and produces docs/audits/ output. |
security-audit | Security audit across auth, input validation, LLM-specific attacks, data access, APIs, secrets, and infrastructure. |
spec | Adversarial spec interview for long autonomous runs, including takt/factory handoff material. |
Productivity
| Skill | Description |
|---|---|
orchestrator | Senior-architect delegation mode for decomposing non-trivial work, routing worker agents, and reviewing output. |
ui | Generate complete HTML+CSS pages in International Typographic Style. |
component | Generate self-contained UI components such as buttons, cards, modals, and tables. |
page | Generate full page layouts such as landing, pricing, dashboard, and blog pages. |
refine | Iterate on generated POCs until the design is approved. |
prd | Convert approved UI POCs into implementation-ready Product Requirements Documents. |
product-context | Guided interview that captures product identity, value proposition, aha moment, and user journey. |
brand-voice | Guided interview that produces voice traits, tone map, customer language inventory, and competitive positioning. |
ux-audit | SaaS UX audit that scans user-facing touchpoints and applies established UX frameworks. |
ux-copy | Writes production-ready UX copy for user-facing surfaces using microcopy and CTA frameworks. |
Misc
| Skill | Description |
|---|---|
client-proposal | Generate client-facing proposals with a built-in review for internal leaks. |
debug | Track a debugging session with a breadcrumb file so work can be resumed without losing context. |
skill-review | Review Claude Code skill files for format, interaction design, chain integration, and effectiveness. |
Reference Docs
Reference docs live in docs/ and are installed by ./install.sh.
| Doc | Used By |
|---|---|
docs/agentic-patterns/patterns-master-summary.md | patterns, agent-review, agent-audit, harness-architect |
docs/agentic-patterns/patterns-selection-guide.md | patterns, agent-review, harness-architect |
docs/agentic-patterns/patterns-anti-patterns.md | patterns, agent-review, agent-audit, harness-architect |
docs/harness-patterns/harness-types.md | harness-architect, harness-review |
docs/harness-patterns/harness-selection-guide.md | harness-architect, harness-review |
docs/skill-authoring-guide.md | harness-build, skill-review, skill authors |
docs/deep-review/framework.md | deep-review |
docs/deep-review/resources.md | deep-review |
docs/deep-review/security.md | deep-review |
docs/deep-review/absence.md | deep-review |
docs/deep-review/research-dimensions.md | deep-review |
docs/skill-stewardship-module.md | Skill stewardship module design and non-goals |
Platforms
Claude Code is the supported platform today. The repo includes a first platform manifest and adapter docs as a starting point for targeting Claude Code, Codex, OpenCode, and Pi-mono.
| File | Purpose |
|---|---|
manifest/platform-matrix.yaml | Declares supported/planned platforms, install targets, artifact types, and retired repo-managed artifacts. |
platforms/claude-code/README.md | Documents the current Claude Code install paths and compatibility rules. |
Modules
Modules are optional infrastructure around the agent/skill system. They are not installed by default. They must stay advisory unless a user explicitly opts into stronger behavior.
| Module | Purpose |
|---|---|
modules/skill-stewardship/ | Advisory skill lifecycle module for suggesting new skills, improvements, splits, redirects, retirements, and later verification. |
platforms/claude-code/skill-stewardship/ | Claude Code hook wiring notes for the Skill Stewardship module. |
Installation
git clone git@github.com:Sebstrdigital/agentic-tools.git
cd agentic-tools
./install.sh
The installer copies:
agents/*.md->~/.claude/agents/docs/*/*.md-> matching~/.claude/docs/<subdir>/docs/*.md->~/.claude/docs/skills/<category>/<skill-name>/*.md->~/.claude/skills/<skill-name>/
It also removes repo-managed legacy command files from ~/.claude/commands/ after their source files migrate to skills. User-owned commands are left untouched.
Safe install behavior:
- If a file with the same name already exists and came from this repo (
source_id: seb-claude-tools), it gets updated with version tracking. - If a file with the same name exists but belongs to something else, it installs with a
sebstrdigital-prefix to avoid conflicts. - If no conflict exists, it installs directly.
After installing, restart Claude Code to pick up changes.
Updating
Pull the latest and re-run the installer:
cd agentic-tools
git pull
./install.sh
The installer shows version changes, for example v1.0.0 -> v1.1.0, and skips files that are already up to date.
File Structure
agentic-tools/
├── agents/
│ ├── builder.md
│ ├── grunt.md
│ ├── heavy.md
│ ├── scout.md
│ └── skeptic.md
├── docs/
├── manifest/
│ └── platform-matrix.yaml
├── modules/
│ └── skill-stewardship/
├── platforms/
│ └── claude-code/
│ ├── README.md
│ └── skill-stewardship/
├── skills/
│ ├── engineering/
│ ├── misc/
│ └── productivity/
├── CLAUDE.md
├── install.sh
└── README.md
Metadata
Repo-managed agent, doc, and skill files include YAML frontmatter with:
source_id: seb-claude-toolsto identify files from this repo.version: 1.0.0to track updates.
The installer uses this metadata to update repo-managed files in place while avoiding unmanaged local files.