Skip to content
Course contents · Get better answers from your AI assistant

Lesson 2 of 5

Write a system prompt that works.

The instructions your assistant reads before every chat: role, context, format, constraints and examples, with a before and after you can copy.

A system prompt is the job description you give a new hire on day one. They read it before they talk to anyone, and it shapes every conversation after. Without it you have a polite generalist. With it you have someone who works for you.

What a system prompt is

A system prompt is the set of instructions an assistant reads before every conversation. The people chatting with it don't see it. Tools call it different things: instructions, custom instructions, the assistant's brief. It's the same idea.

System prompt
Written once, read before every chat. Who it is, who it serves, how it works, what it never does.
Message
Written each time. The question or task for this conversation.

Course one's prompts lesson was about briefing a contractor for one job. A system prompt is the standing brief for a job that repeats. The six questions from that lesson all fit into these five parts: the job and who it's for go in the role, what to use goes in the context, what good looks like goes in the format and the examples, and what to do when unsure and when to stop go in the constraints.

The five parts

01Role
Not "you are a helpful assistant". Instead: "you handle tenant enquiries for a small lettings agency".
02Context
Who the people are, what the business cares about, and how it sounds.
03Format
What a good answer looks like: its structure, its length, what comes first.
04Constraints
The don't-do-this rules: phrases to avoid, topics to stay out of, what to do when unsure.
05Examples
One or two real answers you liked, with "match this tone and structure".

Examples are the step people skip, and often the most useful. A model is very good at continuing a pattern. One real reply shows tone, length and structure better than a paragraph describing them.

Before and after

Same assistant, same question, a tenant asking about a broken boiler. First with no system prompt:

BeforeExample
Dear valued tenant, I hope this email finds you well. Thank you for reaching out regarding your boiler. We understand how inconvenient this must be...

Then with a five-part system prompt:

AfterExample
Hi Jo, sorry about the boiler. I've passed it to our repairs team, and they'll call you to book a time.

Nothing about the model changed. The second answer is short because the format asked for short, says who will act because the format asked for that, and skips "I hope this finds you well" because a constraint banned it.

Two common mistakes

Rules that fight each other. "Be brief" and "cover everything" can't both win, so the model picks one, and not always the same one. If two rules pull in different directions, say which one wins.

One prompt trying to do every job. A prompt for tenant replies and social posts and monthly reports ends up vague about all three. Keep it to one role, and make another assistant for the next job.

Test it on real messages

Write it, then try a handful of real messages covering the main things people ask. Does it stay in role? Follow the format? Sound like you? When it drifts, tighten that one part of the prompt and try the same messages again.

Running example

The running job

Chasing last month's unpaid invoices, one lesson at a time.

Here is Sam's system prompt for the tenant helper, in the five parts. The length limit is Sam's own house rule, not a magic number.

The system promptExample
Role: you handle tenant enquiries for Sam's small lettings agency.
Context: our tenants rent flats and houses we manage. We're friendly and use plain English. We never promise a date we can't keep.
Format: say what happened, what happens next, and who will do it. Under 150 words.
Constraints: no "I hope this finds you well". No legal advice. Never quote a fee. If you don't know something, say so instead of guessing.
Example, match this tone and structure: "Hi Tom, thanks for letting us know about the leak. I've passed it to our repairs team, and they'll call you to arrange a visit. Sam"

Notice the example isn't the boiler reply. It's a different real reply, so the assistant copies the tone and structure rather than the words.

Practice

Try it yourself

About five minutes, in any AI chat app

Write a five-part system prompt

  1. Pick one job you ask an AI to do repeatedly, such as replying to customer emails.

  2. In any AI tool that has custom instructions, or at the start of a new chat, write the five parts: role, context, format, constraints and one real example of yours.

  3. Ask the same question you'd normally ask, with and without the prompt, and compare.

What to look for

The answer with the prompt should need fewer edits before you'd send it. If one part still comes out wrong, the matching part of the prompt is the one to tighten.

Common mistakes

  • "You are a helpful assistant" as the whole role.

    Name the job and who it's for: "you handle tenant enquiries for a small lettings agency".

  • Describing the tone in adjectives and giving no example.

    Paste one real answer you liked and add "match this tone and structure".

  • Rules that contradict each other.

    Read the constraints side by side and say which one wins when they clash.

Quick check

3 questions to see what stuck. Nothing is scored or saved.

  1. 1. Who sees the system prompt?

  2. 2. Which of the five parts do people most often skip?

  3. 3. Your prompt says "be brief" and "cover everything". What's the problem?

Remember

A system prompt is the day-one job description: role, context, format, constraints, and a real example to copy.

Sources and further reading

  1. 1.The blog post this lesson grew out of: How to write AI system prompts that actually work

Optional

Try it in LaunchLemonade

Opens a new assistant with Sam's five-part prompt from this lesson already in its instructions. Swap in your own business and an example of yours. Everything above works in any AI tool.

Open this prompt as an assistant

Finished this lesson?

0 of 5 done in this course. Progress stays in this browser, no account needed.