System Prompt

You already hired the intern. The system prompt is the standing orders they read before you walk in.

Tonight's message is the job. The standing orders are who they are, what they do, and what they never do — every turn, whether you remember or not. You don't see that paragraph. They do. It shapes every reply.

Same model, two handbooks, two employees. "Be careful, never invent an API." "Be funny, keep it moving." The engine did not change. The orders did.

Three layers, one slot

1. Who to be. Tone. Role.
2. What to do. Stay on this product. Scaffold through the tools, don't hand-roll a stack.
3. What never to do. Don't guess a price. Don't dump a secret. Those are guardrails written as words.

CLAUDE.md and AGENTS.md are house rules on disk. The harness copies them into this same standing slot at the start of a session. A skill is a recipe you pull for one job. The system prompt is the recipe that is always on.

This repo's plugin does not replace that slot. It fills a shelf next to it: skills, slash commands, MCP tools. The host still decides the standing voice. You still decide whether the intern invents a kitchen or calls npx create-hackhaton-space-stack@latest mcp.

Keep it short. A novel of rules is a novel they will skim. When the agent keeps doing the thing you hate, stop arguing in the thread. Change the orders they read every morning.

What this unlocks

Once you see the standing slot, "the model is being weird" stops being a personality. You are either writing the handbook, or you are yelling at someone who already clocked in with yesterday's.

A prompt injection tries to overwrite that handbook from a webpage. Your job is to keep the handbook yours.

Talk to the intern after you write the orders. Not instead.