Agents MD Snippets

AGENTS.md Snippets

Copy-paste conventions from my AGENTS.md / CLAUDE.md files, as shown in my videos on coding with AI agents:

- The 4 Levels of Coding With AI
- The live multi-agent session (link coming when it's published)

These work in any agent instruction file — AGENTS.md, CLAUDE.md, or equivalent. Adjust the wording to your own setup; the ideas are what matter.

1. Context re-entry (the context-switching fix)

When you're juggling several agents at once, every summary an agent hands you is a cold start for *you*. This block makes agents write for that. It lives in my root instruction file, shared across all repositories.

2. Git worktrees for parallel agents

3. TDD is mandatory

4. Builder/driver split for the no-mistakes gate (the token optimization)

This is the optimization from the live session: the agent that *built* a feature is sitting on a huge context (often 150k–200k tokens). If that same agent drives the review gate — which takes many turns of monitoring and simple responses — that entire context is resent on every turn. Instead, the builder hands off, and a fresh, cheap, tiny-context agent drives the gate. A park→decide→resume roundtrip costs ~30k tokens on a driver vs ~200k on a builder.

Adapt to your setup; mine targets the no-mistakes gate.

Bonus — cross-provider review: having a different provider review than the one that built (e.g. Claude Code implements, Codex reviews, or vice versa) catches a different distribution of bugs, and spreads the token load across two subscriptions.

Tools mentioned in the videos

- no-mistakes gate by Kun Chen — automated review, tests, and screenshot evidence before code reaches you
- Grill Me skill by Matt Pocock — makes the agent interview you through every design decision, one question at a time
- Wispr…