How to Give an AI Agent Context About You (Before It Acts for You)

For the last few years, AI mostly talked. You asked, it answered, and you decided what to do with the answer.

Agents change that. An AI agent doesn't stop at a suggestion. It books the meeting, sends the email, fills the cart, or edits the file. Anthropic's engineering team draws the line cleanly in its guide to building effective agents: agents are "systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

That's a real shift in what context is for. When a chat assistant doesn't know you, you get a generic answer and lose thirty seconds. When an agent doesn't know you, it does the wrong thing, and it does it in your name.

So here's what an agent needs to know about you before it acts, and how to write it down so it actually gets followed.

How do you give an AI agent context about you?

Write one short document the agent loads before every task, covering four things: who you are, your standing rules, your hard limits, and what's current. Keep it specific and under a couple of pages. Back the hard limits with the tool's own permission settings, and refresh the current section every few weeks.

	Chat assistant	AI agent
What it does	Answers a question	Takes an action for you
Cost of missing context	A generic reply	A wrong email, booking, or purchase
Who catches the mistake	You, before anything happens	Often the other person, after it's done
What it needs most	Background and voice	Background, voice, rules, and limits
Where the context lives	A prompt or custom instructions	A file or instruction set it loads every time

Why does an AI agent need more context than a chatbot?

Because the safety net moved.

With a chatbot, you're the last step. Every answer passes through your judgment before it touches the world. A bad draft just sits there until you fix it or throw it away.

With an agent, that step shrinks or disappears. If it doesn't know you never take calls before ten, it accepts the 8 a.m. invite. If it doesn't know a client prefers formal email, it sends the breezy one. Nobody reviewed it. Reviewing it was the job you handed off.

The thing is, a modern agent is plenty capable. What it lacks is you. A capable model with no context about your preferences will make the reasonable choice for an average person, and the average person isn't you. That gap was tolerable when the output was text. It's expensive when the output is an action.

What does an AI agent need to know about you?

Four things. Each one does a different job.

1. Who you are and how you like things done. Your role, who you work with, and the texture of how you operate. How formal your emails are. How fast you reply. Whether you prefer one long meeting or three short ones. This is the same background you'd give any assistant, and it's the part most people already have scattered across chats and settings.

2. Standing rules. The corrections you'd otherwise make every single time. "Meetings only Tuesday through Thursday, 10 to 4." "Sign emails with my first name only." "Always copy my business partner on anything over $500." If you've said it twice, it belongs here as one short rule.

3. Hard limits. The actions the agent must never take without asking. This is the section chat assistants never needed and agents can't do without. Spending money. Emailing someone new. Deleting or overwriting anything. Sharing client details. Accepting on your behalf. Write each one plainly. No nuance, no "usually."

4. What's current. Active projects, live deadlines, and the people in play this month. An agent acting on last quarter's priorities is worse than one with no priorities at all, because it acts with full confidence. Stale context is the quiet failure here, and it's the one most worth scheduling time against.

How do you write agent context that actually gets followed?

Short, specific, and organized. That comes straight from how the tools describe themselves.

Anthropic's documentation for Claude Code, one of the more widely used agents, is unusually direct about this. On how its instruction files (CLAUDE.md) get used, it says: "The more specific and concise your instructions, the more consistently Claude follows them." It recommends keeping each file under 200 lines, because "longer files consume more context and reduce adherence," and grouping instructions "under markdown headers and bullets." It also warns that "if two instructions contradict each other, Claude may pick one arbitrarily."

Those are developer docs, but the lessons travel to any agent.

Replace wishes with rules. "Be careful with my calendar" gives an agent nothing to act on. "Never accept a meeting that overlaps a focus block" does.

Cut anything that wouldn't change what the agent does.

Hunt down contradictions. If one line says "reply to every client within an hour" and another says "never email after 6 p.m.," the agent has to guess which one wins.

And then the most important line in that documentation: Claude Code treats these files "as context, not enforced configuration." Written rules guide an agent. They don't physically stop it. For anything you truly can't afford, like purchases or deletions, use the tool's own approval or permission settings too. Belt and braces.

What does it look like in practice?

Take a hypothetical freelance garden designer who hands an agent her inbox and calendar.

Without context, the agent is polite, efficient, and slightly wrong all day. It confirms a site visit during school pickup. It accepts a call from a vendor she's been avoiding.

With one page of context, the same agent behaves like someone who's worked with her for a year. The who-you-are section says she's a one-person studio, writes warmly but briefly, and signs with her first name. Standing rules say site visits go on Monday and Wednesday mornings, and suppliers get short, direct replies. Hard limits say never accept anything from someone new, never agree to a price, and never send a design file without asking. The current section names her two active projects and the one client deadline this month.

That page takes an evening to write. It saves her from the kind of mistake that takes a week of apologizing to undo.

Can you use the same context for chat and agents?

Mostly, yes, and you should.

Who you are, how you write, and what you value don't change when the tool starts taking actions. That's the durable layer, and it should live in one place you own, not locked inside one vendor's settings. Keep it portable and it moves between ChatGPT, Claude, and Gemini as easily as it moves into an agent's instructions.

What an agent adds is a stricter layer on top: the standing rules and hard limits. Think of the shared background as the person, and the agent rules as the job description. This is context engineering at the scale of one person, deciding on purpose what an AI has in front of it before it does anything.

Where to start

Start with the durable layer, because every agent you use from here on will need it.

RUMO breaks that layer into six short documents, so your values, voice, stories, current situation, timeline, and the people in your work each live in their own file, and you update the one that changed without rewriting the rest.

The first one is values: what you care about, what you won't do, and how you make calls. That's the part an agent can't infer from your inbox, and it's where good hard limits come from. The free Personal Constitution builder takes about thirty minutes.

AI used to wait for you to act on its answers. Now it acts. The least you can do is tell it who it's acting for.