convez.ai
MCP ServerProdutividadeIntermediário
CodyAMaughan

Agent-First Engineering

CodyAMaughan/agent-first-engineering

Curriculum + scaffolder para arquitetar repos onde agentes de IA prosperam

Ver no GitHub
BomPluginScore · de 0 a 100

Currículo e scaffolder para desenhar codebases otimizadas para agentes de IA (Claude Code, Codex, Cursor). Fornece uma metodologia de 6 fases com padrões abertos (AGENTS.md, SKILL.md) e um gerador interativo que configura novos repositórios ou refatora existentes para sucesso com agentes.

Instalação

$npx convez-ai CodyAMaughan/agent-first-engineering

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
3
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 3 meses
Versão
v0.1.0
Issues abertas
1
Contribuidores
1

Casos de uso

  • Estruturar projetos para máxima eficácia com Claude Code e Cursor
  • Treinar times em eng. otimizada para agentes (6 fases didáticas)
  • Gerar AGENTS.md e SKILL.md com entrevista interativa no agent

Veredito

Obra madura e bem estruturada que leva a sério a engenharia com agentes, indo além de "vibe coding". O currículo é clara e o scaffolder resolve um problema real: padrões agent-agnostic reutilizáveis. Recomendado para equipes que já usam agentes e querem desenho técnico rigoroso.

README

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

Stop coding with AI agents "loosey-goosey." Design your codebase so agents succeed — the way a systems engineer would.

📖 Read it online → codyamaughan.github.io/agent-first-engineering

This repository has two deliverables that are two views of one body of knowledge:

  • A — The Curriculum (specs/001-curriculum/) — a phased, visual course that takes you from informal "vibe coding" to designing agent-first codebases. Modeled structurally on the well-organized AI Engineering from Scratch approach: one concept at a time, diagrams, ELI5, a real artifact at the end of every lesson.
  • B — The Scaffolder (specs/002-scaffolder/) — an agent-agnostic, SKILL.md-first tool that interviews you about a project, then generates a proper agent-first setup: AGENTS.md, a SKILL.md library, and lifecycle-hook guardrails — wired to work across Claude Code, Codex, and Cursor (more agents via adapters). Think create-react-app for agent-first repos, but a conversation instead of a fixed form, and it runs inside your agent. Modes: init (new repo), adopt (clean up an existing one).

Teach and generate in lockstep: every layer the curriculum teaches, the scaffolder generates; every artifact the scaffolder generates, the curriculum explains.

Who this is for

  • Developers already using AI agents who want to level up from "vibe coding" to designing repos where agents succeed by default.
  • Tech leads deciding how their team should adopt AI coding tools.

It assumes you've used an AI coding agent before. It is not an intro-to-AI course.

What you'll get

  • A 6-phase course (read it free at the site) — from prompting and context engineering through verification, memory, spec-driven development, and harness engineering.
  • A scaffolder that interviews you and sets up a new (or existing) repo to be agent-ready — AGENTS.md, a skills library, and guardrail hooks — for Claude Code, Codex, and Cursor.

Principles

The project is governed by its Constitution. In brief:

  1. Open Standards First — AGENTS.md + SKILL.md are the source of truth; vendor formats are optional adapters.
  2. Agent-Agnostic by Construction — author once, render per-agent. Claude Code is the reference, not the requirement.
  3. Teach and Generate in Lockstep — A and B stay in sync, by rule.
  4. Guardrails Over Vibes — correctness is enforced by hooks/tests/CI, not prose.
  5. Minimal Context, Progressive Disclosure — short, command-first, machine-parseable.
  6. Adopt, Don't Reinvent — build alongside mature, permissive tools (esp. GitHub Spec Kit), don't fork them.
  7. Specs Are the Source of Truth — spec → plan → tasks → implement.

What we build on (all permissive, all current)

LayerAdopted standard / toolLicense
Context fileAGENTS.mdOpen standard
Reusable skillsAgent Skills / SKILL.mdApache-2.0
Spec workflowGitHub Spec Kit (complement, not fork)MIT
Principles12-factor-agentsApache-2.0

See meta/prior-art.md for the full landscape and why this project's niche is currently unfilled.

The Curriculum

Six phases, from vibe coding to systems engineer for agents. Full index in docs/curriculum/. Quiz yourself with /check-understanding <phase> (the check-understanding skill generates an interactive quiz from each phase's lessons).

#Phase#Phase
1Fundamentals4Session & Memory
2Context Engineering ★★★5Spec-Driven Development
3Verification & TDD ★★★6Orchestration & Harness

Project status

Early and active. The foundations tier is complete and published; the Advanced Patterns tier is on the Roadmap. Built (and dogfooded) using GitHub Spec Kit.

FAQ

No. It's agent-agnostic — Claude Code is the reference implementation, not a requirement. Lessons show the open-standard form plus Codex and Cursor equivalents.

Yes — MIT-licensed and open source.

No, and I'm upfront about it: it's a curation and synthesis of the best blogs, posts, and documentation from the top AI labs, their contributors, and other thought leaders — cited throughout. It was drafted primarily with AI (fitting, given the topic) and reviewed by me. What's mine is the direction, the structure, and the review.

The curriculum teaches the practice; the scaffolder generates it. Each lesson ends by showing what the scaffolder produces.

See Contributing — fixes, diagrams, better sources, and new topics from the Roadmap are all welcome.

Contributing

Contributions are welcome — fix a lesson, add a diagram, propose a topic from the Roadmap, or improve the scaffolder. Start with CONTRIBUTING.md and our Code of Conduct.

License

MIT © 2026 CodyAMaughan. Redistributed dependencies are MIT / Apache-2.0 / BSD.

Ver o original no GitHub

Bom para

Relacionados