Skip to content

Agent-First Engineering

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

Stop coding with AI agents loosely. Design the repository and its surrounding harness so that any competent agent succeeds by construction — then the human's job shifts from reviewing every diff to engineering the environment the agent operates in. The bet isn't on a vendor; it's on the open standards the major coding agents converged on in 2026 (AGENTS.md, SKILL.md, MCP) and on two load-bearing disciplines: context engineering and verification. That makes agent-first setup a portable, teachable discipline rather than a vendor lock-in bet.

This project is two views of one body of knowledge:

  • The Curriculum — a phased, visual course that takes you from informal "vibe coding" to designing agent-first codebases: one concept at a time, diagrams, ELI5, a real artifact at the end of every lesson.
  • The 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.

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

Start here

  • 🎓 The Curriculum → — the full 6-phase course, from vibe coding to systems engineer for agents. Start with Phase 1 — Fundamentals.
  • 🚀 Roadmap → — the Advanced Patterns tier coming next (skills/hooks deep-dives, security, token optimization, cloud/voice/mobile agents, and more).
  • 🗺️ Translation Matrix → — deep Claude→Codex→Cursor feature research, by layer.

What we build on

All permissive, all current open standards:

Layer Adopted standard / tool
Context file AGENTS.md
Reusable skills Agent Skills / SKILL.md
Spec workflow GitHub Spec Kit (complement, not fork)
Principles 12-factor-agents

The two load-bearing competencies — context engineering (treat the window as the scarce resource) and verification (give the agent an external oracle so it closes its own loop) — carry the most weight across every researched source. Everything else is scaffolding around those two.