Given at Coimbra.js 2026 ↗
The failure mode I kept seeing: an agent with full context in the chat, nothing written down where it could find it on its own. Every new task started with ten minutes of re-briefing that should have been a file.
The talk covers six primitives: CLAUDE.md for persistent context, Skills for repeatable procedures, Slash Commands for common operations, Subagents for parallel work, Hooks for automated guardrails, and Plugins for extending the agent's reach. The goal is a project an agent can pick up the same way a good README helps a new hire.
The examples are from my homelab. Agentic development is a project hygiene problem more than a prompt engineering one.