Instructions: Your Agent's Brain

What it is#

An Eve agent begins as a folder. The single most important file is instructions.md — a Markdown document that tells the model who it is and how it should behave. It is the always-on system prompt.

Why you'd care#

You can get a talking agent with zero TypeScript. Instructions are where you set tone, scope, and guardrails before any code exists.

Before#

An empty directory. No agent, no behavior.

After#

Scaffold the project and write your first instructions:

npx eve@latest init my-agent
cd my-agent

# Identity
You are a personal research assistant. You help the user understand
topics by giving grounded, citation-friendly answers.

# Guardrails
- Never invent facts. If you don't know, say so.
- Prefer primary sources and clearly separate facts from opinion.

Run it:

eve dev

Then ask: "What's the difference between durable and ephemeral execution?" — the agent answers from its instructions.

Do it yourself#

1. Run npx eve@latest init my-agent and cd my-agent.
2. Open agent/instructions.md and set an identity plus two guardrails.
3. Run eve dev and hold a short conversation.
4. Confirm the agent stays in character and respects your guardrails.

Gotchas#

• instructions.md is always in context — keep it specific, not a novel.
• The filename matters: instructions.md is the reserved name Eve discovers automatically.

Recap#

instructions.md is your agent's brain. One file, one npx eve init, and you have a talking agent. Next, we choose its model.