Designing the bot's brief

What a system prompt actually does

A system prompt is the standing instruction you give a language model before any user message. It defines who the model is in this conversation — what it knows, what it cares about, what its tone is, what it must do, and what it must refuse.

You used Claude or ChatGPT in the first course without a system prompt. You just chatted. The model was being its general self. For a research bot, that is not enough. The general self of a language model will happily answer "should I buy this stock", give you a confident-sounding take, and probably suggest you "consider your risk tolerance". That is not the bot you want.

The bot you want is a specialist — focused, disciplined, sourced, and prepared to say no to questions it should not answer. The system prompt is how you turn the general model into that specialist.

The six things a research-bot system prompt must contain

A serviceable system prompt for this kind of bot has six parts.

1. Role. Tell the model what it is. Not "you are an AI assistant", but specifically: "You are a market and company research assistant. Your job is to help the user understand companies and markets by reading and summarising material the user provides."

2. Scope. Tell the model what it does. Read, summarise, compare, characterise sentiment, answer factual questions, surface what a document says — and only that.

3. Refusal. Tell the model what it must refuse. This is the most important part. "You will not give buy/sell recommendations, predictions about price movements, or any form of investment advice. If asked, you will explain that this is not what you do."

4. Sourcing. Tell the model how it must present its claims. "Every factual claim you make must be sourced to a specific part of the material the user provided. If you cannot source a claim, say so. Never invent figures, dates, or quotations."

5. Tone. Tell the model how to sound. "Plain, precise, professional. No hype. No 'in conclusion' filler. If something is uncertain, say it is uncertain. If something is unknowable, say it is unknowable."

6. Calibrated confidence. Tell the model how to handle ambiguity. "Distinguish what the document says from what you infer from it. Distinguish a fact from an interpretation. Never assert with confidence something you are guessing at."

A worked example: the full system prompt

Here is a system prompt that includes all six things. You are welcome to copy this exactly, or use it as a starting point and adapt to your needs.

That is the entire prompt. It is about 350 words. It does a lot of work.

Why the refusal language matters so much

Pay attention to the wording of the refusal section. Three things make it robust.

First, it lists the prohibited things specifically. "Buy/sell recommendation", "price prediction", "portfolio allocation". The model is much more likely to refuse a specific request when the prompt named that specific request than when it just said "no investment advice".

Second, it provides an alternative. "Instead of refusing flatly, offer to help research the underlying question." A pure refusal often pushes users to phrase the question differently and try again. An alternative gives them something useful to do that does not require breaking the rule.

Third, it includes the line "You will not be talked out of this rule." This sounds dramatic, but it works. Without it, a clever user can sometimes argue the model out of its constraints — "but I'm just curious", "but it's hypothetical", "but my friend already decided". With it, the model has a clear instruction to hold the line.

Hands-on time

What you should have seen

On the first run, the strong models — Claude 3+ and GPT-4+ — usually hold the refusal cleanly on all three stress tests. Smaller or older models sometimes slip on the hypothetical or the "your honest opinion" framing.

If the refusal slipped, look at the exact wording. Often the model produced something that looks like advice but technically isn't — "the stock has had a strong year and many analysts are bullish, but past performance is not a guarantee". This is a near-miss. Strengthen the system prompt by adding: "You must not state whether analysts, the market, or anyone else is bullish, bearish, optimistic, or pessimistic about a specific security. You may report what specific named sources have said, with citations, but you may not summarise market sentiment as a recommendation in disguise."

Save your prompt

The system prompt you have just built is the most reusable thing in this course. Save it somewhere you can paste it again — at the start of every new research session you have with the bot. For the rest of the course, we will assume your bot has this prompt loaded.

Looking ahead

In Lesson 3 we give the bot real material to read. You will paste in an actual annual report or earnings transcript and watch the bot work it. This is where the bot starts being useful.