Instructions

Strip everything else away and an agent is just a model in a loop. You give it a prompt, it gives you text back. That's already useful — it can explain a concept, draft a plan, write an email. But it has an obvious boundary: the model only sees what's inside the prompt, plus whatever it was trained on. If your project, your ticket, your incident isn't in that input, the model starts guessing. We call that hallucination.

So the first question is never "is the model smart enough?" It's "what are we giving it to work with?"

Instructions are the first answer. You tell the model who it is, what kind of work it's doing, the tone, the constraints, the coding style, the rules for reviewing code, the things it must never do. That standing guidance is the first harness primitive.

Guidance moves into the environment

The power of instructions is that they relocate repeated guidance out of your head and into the system. Without them, every session starts with you re-typing "keep answers short, follow our lint rules, never commit without asking." With them, the harness injects that guidance automatically — you never paste it again.

# CLAUDE.md — loaded into every session, automatically

- This is a Meteor.js app. The app root is `meteor/league/`.
- Normal work stays inside `imports/api/` and `imports/ui/`.
- Run the focused test suite before claiming a fix works.
- Never touch migrations without asking first.

In the wild

Every serious harness has this layer, under different names:

• CLAUDE.md in Claude Code — project instructions loaded at session start
• AGENTS.md — the cross-tool convention supported by Codex and others
• System prompts and custom instructions in raw API use
• Cursor rules, repository rules, org-level policy files

Different names, one primitive: durable instructions the harness delivers so you don't have to.

Instructions shape everything — and discover nothing

Instructions are the first harness layer because they shape the behavior of everything that happens with the model afterwards. Tone, caution, conventions — all downstream of this file.

That gap — between telling the model how to act and showing it what to act on — is exactly where the next primitive begins.