Steering

The system prompt is how you steer an agent.

The model is a generic engine. The system prompt is where you give it an identity, a job, and its rules. It is the steering wheel.


Swap nothing but the system prompt and the same model becomes a careful code reviewer, a terse data extractor, or a patient tutor. Almost everything you want an agent to be, you say here.

What it does

The text that rides along on every turn

The system prompt sits at the very top of every context window, ahead of the conversation. It tells the model who it is, what it is allowed to do, what it must never do, how to go about its work, and the shape its answers should take.

Because it rides along on every single turn, it is the most leveraged text in the system. A one-off instruction can be forgotten as the conversation scrolls past; the system prompt is always present, so it shapes behavior far more reliably.

system prompt present on every turn
Role who the agent is
Capabilities what it is allowed to do
Constraints what it must never do
Workflow how it works, step by step
Output format the shape its replies should take
Because it rides along on every single turn, the system prompt is the most leveraged text in the whole system.

a system prompt, filled in

# role

You are a release-notes writer for a software team.

# capabilities

You may read the diff and the changelog. You cannot edit code.

# constraints

Never invent a feature that is not in the diff.

Never include internal ticket numbers.

# workflow

1. Group changes into Features, Fixes, and Chores.

2. Write one plain line per change, newest first.

# output format

A single markdown list.

The same five parts from the diagram above, written out: specific, ordered, with the constraints stated as hard rules.

Writing one that works

What makes a good system prompt

The structure is the easy part. Making an agent actually behave takes a handful of habits that hold up in practice.

Techniques for writing an effective system prompt
Technique Why it works
Order it: role and constraints first Rules stated up front are weighted more heavily and are harder to miss than ones buried mid-prompt.
Be specific, not verbose A short, concrete rule is followed more reliably than a long, hedged paragraph.
Use hard gates "Stop, you must do X before Y" is obeyed where "please try to" is quietly skipped.
Show an example One example of the output you want shapes the model more than a sentence describing it.
Name the steps For multi-step work, label the phases so the agent can track where it is.

Why it is the strongest lever

Where intent lives

It is worth knowing what the prompt cannot do. It cannot grant an ability the tools do not provide, and a model can still drift from even a well-written one over a long task. Words steer; they do not enforce.

That is why steering is paired with harder controls: the gates and limits of oversight. The model supplies the raw capability, the prompt supplies the purpose, and it is still the cheapest, highest-leverage change you can make, and usually the first place to look when an agent misbehaves. Just not the only one.

Tools give an agent hands and the loop gives it persistence, but the system prompt gives it intent. It is the difference between a capable engine and an agent that does the thing you actually wanted.

Next: oversight, the harder controls How the prompt is assembled into context Back to the overview