You've reached the end of the track — and the course. Now is not the time to learn more concepts: it's time to build your system. In this module, you'll build your personal subagent library, start from a ready-made starter pack, learn templates that save hours, and review in a screen everything the course taught. At the end, a final project: create your own agent suite from scratch.
Detailed content
🗄️ Your library of specialists
Agents are files. Files can be organized. Organization becomes power.
What it is
A library of specialists is a set of files .md with subagent definitions that you maintain and evolve. Each file is an “employee” with a clearly defined name, scope, tools, and restrictions. The library lives in two places: ~/.claude/agents/ (available in all projects) and .claude/agents/ within each project (repository-specific).
Why build your library
- ▸Avoids rewriting the same system prompt every time—call the agent by name
- ▸Well-defined agents make fewer scope errors — they know what they should not do
- ▸Improves over time: fix a bad agent once and every project benefits
- ▸Teams can share the library via git — an emerging organizational standard
Key concepts
📦 Starter pack — 5 agents to get started today
security-auditor · test-runner · doc-writer · db-expert · plan-roaster
What it is
A set of five agents covering the most common use cases in software development. They’re generic enough for any stack, but specific enough to be useful without customization. Copy them, adjust the details for your project (database names, frameworks, testing conventions), and you have a functional suite in minutes.
Why these five
Key concepts
- ✓Minimal tools (only what's needed)
- ✓Explicit restrictions (“never execute X”)
- ✓Well-defined scope per agent
- ✓Output format specified
- ✗Agents that do everything (no focus)
- ✗Writing permissions without a reason
- ✗System prompts over 200 lines
- ✗Circular dependencies between agents
📐 Ready-made templates for your suite
Frontmatter + body in minutes, not hours.
What it is
An agent template has three parts: the YAML frontmatter (metadata that Claude Code reads), the system prompt (how the agent should behave) and the usage examples (optional but very useful). A good template is a shortcut for creating new agents — fill in the variable parts and you have a working agent in 5 minutes.
Why use templates
- ▸Consistency: all agents in the suite have the same output structure
- ▸Speed: new agent in 5 min without thinking about the structure
- ▸Onboarding: a new contributor understands the convention without extra documentation
Key concepts — template anatomy
✅ Production checklist
Before putting an agent into real-world use, go through this list.
What it is
A checklist is the last filter before an agent becomes part of your permanent library. It covers security (minimum tools, no unnecessary permissions), quality (clear system prompt, defined output format), and operations (who calls it, when, what happens if it fails).
Why use a checklist
Bash without restrictions, subprocesses can create loops or consume resources unexpectedly. Always test in an isolated environment first.Key concepts — the concise checklist
🔀 Project × Global in practice
What goes to ~/.claude/ and what goes in .claude/?
What it is
Claude Code resolves agents in two scopes: global (~/.claude/agents/) and local (.claude/agents/ in the project). If an agent exists in both places with the same name, the local one overrides the global one — allowing you to "override" generic specialists with project-adapted versions without removing the global one.
Why this matters in practice
- • generic security-auditor
- • universal doc-writer
- • default plan-roaster
- • Agents for constant personal use
- • project stack db-expert
- • test-runner with the right framework
- • style-reviewer for the team guide
- • Domain-specific agents
Key concepts
🧠 Overall recap — the course on one screen
Everything that matters, in one-line sentences.
What it is
This is slide 18—the course's closing screen. Each rule below summarizes an entire concept you studied. If one of them sounds unfamiliar, that's a sign of which module to revisit. If they all made immediate sense, you finished the course with the right mental map.
🤖 Subagents: what stays
.claude/agents/
~/.claude/agents/
— end of course · INEMA.CLUB/subagentes —
Why a recap on one screen works
- ▸Compact mind map — when in doubt on the job, reread the 8 rules
- ▸Each rule points to course modules for further study
- ▸Makes it easy to share what you learned with a colleague in minutes
🚀 Next steps
What to do after the course ends.
What it is
The course gave you the map. The next steps are the territory. Real practice with subagents happens in real projects — with real mistakes, iterations, and discoveries that no course can simulate. This section points you toward where to go from here.
Why acting immediately matters
Subagent knowledge decays quickly if it isn’t applied. The best window for assimilation is the first 48h after the course. Use that momentum.
Key concepts — where to continue
📄 Real example — a package of 5 agents (.md)
Five files ready to copy to ~/.claude/agents/ or .claude/agents/:
# ~/.claude/agents/security-auditor.md --- name: security-auditor description: Audita código em busca de vulnerabilidades de segurança. Use para revisar PRs críticos, módulos de autenticação, inputs de usuário. NÃO use para refatoração. tools: - Read - Bash # somente leitura — sem escrita de arquivo disallowedTools: - Write - Edit --- # Security Auditor Você é um especialista em segurança de aplicações. Sua única função é revisar código e apontar vulnerabilidades. ## O que você faz - Identifica SQL injection, XSS, CSRF, secrets expostos, dependências com CVE - Classifica cada achado: CRÍTICO / ALTO / MÉDIO / BAIXO - Sugere correção específica para cada achado ## O que você NÃO faz - Não modifica arquivos - Não sugere refatoração fora do escopo de segurança - Não executa código de produção ## Formato de saída ```markdown ## Relatório de Segurança — [arquivo/PR] ### Achados | Severidade | Linha | Descrição | Correção sugerida | ```
# ~/.claude/agents/test-runner.md --- name: test-runner description: Roda suítes de teste, interpreta falhas e sugere correções. Ideal para CI debug e TDD. Use ANTES de mergear. tools: - Bash - Read disallowedTools: - Write - Edit --- # Test Runner Roda testes, lê a saída e explica cada falha em linguagem humana. Sugere correção mas não aplica. ## Regras - Rode sempre com `--no-coverage` para velocidade na primeira passagem - Agrupe falhas por módulo (não por arquivo) - Priorize: build breaks > falhas de lógica > warnings
# ~/.claude/agents/doc-writer.md --- name: doc-writer description: Gera documentação técnica (JSDoc, docstrings, README, CHANGELOG). Use após implementação. NÃO faz refatoração. tools: - Read - Write # só para arquivos de documentação (.md, comentários) --- # Doc Writer Especialista em documentação técnica. Lê código e gera docs claros, concisos e atualizados. ## Convenções - JSDoc para funções públicas - Docstrings para funções com lógica não-óbvia - README: instalação, uso, exemplos mínimos - CHANGELOG: formato Keep a Changelog
# .claude/agents/db-expert.md (local — adapte ao projeto) --- name: db-expert description: Especialista em banco de dados do projeto. Revisa queries, migrations e índices. NÃO executa DDL/DML. tools: - Read - Bash # apenas queries SELECT em banco de dev disallowedTools: - Write --- # DB Expert — [nome do projeto] Stack: PostgreSQL 15, Prisma ORM, ambiente dev em localhost:5432. ## O que faço - Analiso N+1, missing indexes, cartesian products - Reviso migrations antes do deploy - Explico planos de execução (EXPLAIN ANALYZE) ## Formato de resposta Para cada problema: impacto estimado + query corrigida + índice sugerido.
# ~/.claude/agents/plan-roaster.md --- name: plan-roaster description: Critiques implementation plans with constructive skepticism. Use before any complex task. Does NOT modify code. tools: - Read disallowedTools: - Write - Edit - Bash --- # Plan Roaster You are the technical devil's advocate. You read plans and point out what will go wrong. ## How you work 1. Read the complete plan 2. List implicit assumptions (the most dangerous ones) 3. Point out failure points: edge cases, external dependencies, difficult rollback 4. Suggest questions the author should answer before starting ## Tone Direct and useful. Not pessimistic by nature — just realistic.
💬 Ready-to-use prompts
Analyze this project and create 5 custom subagents for me in .claude/agents/. For each one: identify the most-needed role, define the minimum tools and explicit restrictions. Format: one code block per .md file, ready to copy. Start with the most impactful.
Leia o arquivo .claude/agents/[nome].md e aplique o checklist de produção: ferramentas são mínimas? restrições explícitas? escopo claro em 30s? formato de saída definido? Aponte problemas e sugira versão corrigida.
Usando os subagentes disponíveis (security-auditor, test-runner, doc-writer, db-expert, plan-roaster), monte um workflow para revisão completa antes de PR: rode os agentes relevantes em paralelo, colete resultados e gere um relatório unificado com ações prioritárias. Comece pelo plano antes de executar.
Final Project
Design and create your own subagent suite
What to do
- Choose a real project you have or are developing.
- Identify the 3 recurring bottlenecks in your workflow (examples: "I review security manually," "I test by hand," "it takes me too long to document").
- For each bottleneck, design a subagent: name, tools, constraints, output format.
- Create the files
.mdin.claude/agents/from the project. - Test each agent on a real-world case and document what worked and what you adjusted.
- Write one paragraph: "My suite does X. It used to take me Y. Now it takes me Z."
✓ Completion criteria
- ✓≥ 3 .md files created and working
- ✓Each agent passed the production checklist
- ✓Each agent tested on a real case (not simulated)
- ✓An honest, well-written impact paragraph
- ✓At least 1 agent committed to the repository
✗ The exercise doesn’t count if
- ✗The agents were only imagined, not created
- ✗None were tested with a real task
- ✗The "project" is a hello world created just for the exercise
- ✗The impact paragraph was generated by AI without an honest review
You've completed the course Subagents!
You went from “what is a subagent” to building your own factory of specialists. This is one of the most valuable skill sets in AI development in 2026—and you have the complete roadmap.
"A well-built subagent is a colleague who never forgets, never gets tired, and never exceeds their scope."
— The course’s central principle