Don’t dump SDK docs into your Agents.md

Google published codegen instructions for their GenAI SDK here:

googleapis/js-genai codegen_instructions.md

This is gold. It tells you the correct imports, the right model names, how to stream, how to do function calling — everything your agent needs to write correct GenAI code.

But here’s the mistake people make: they copy-paste the whole thing into Agents.md or their system prompt.

Why that’s wrong

• Token waste — you’re burning context on instructions the model only needs when writing GenAI code
• Noise — Agents.md should describe how your agent behaves, not SDK reference docs
• Staleness — SDK docs change, your Agents.md won’t keep up

What to do instead

Use prompt references. Keep SDK instructions in a separate file like docs/gemini-sdk.md and reference it only when the task needs it.

docs/
  gemini-sdk.md       ← SDK codegen instructions live here
AGENTS.md             ← agent behavior, personality, guardrails

Your harness should pull docs/gemini-sdk.md into context only when the agent is about to write GenAI code. Not on every turn.

The pattern

1. AGENTS.md → who the agent is, what it can do, rules of engagement
2. docs/*.md → reference material, injected on demand
3. Harness → the thing that decides when to inject what

That’s how you build a harness that doesn’t choke on its own context. Keep instructions close, but only pull them in when needed.