How to write a system prompt for a website assistant

The single biggest lever on how good your assistant feels isn't the model — it's the system prompt. That's the hidden set of instructions the assistant reads before every conversation. Get it right and a modest model feels sharp; get it wrong and the best model in the world sounds like a lost intern.

Start with who it is and who it serves

Open the prompt by naming the assistant, the business, and the visitor. Something like: "You are the assistant for Acme Bakery. You help customers with orders, hours, and product questions." That one sentence anchors everything that follows. Vague prompts produce vague assistants.

Give it the facts, not a personality dump

People overspend words on tone and underspend on information. The assistant can't answer "do you have gluten-free bread" unless you tell it. Put the real, specific facts in — hours, delivery areas, return policy, the products you actually sell. Keep it current; stale facts are worse than none.

A concrete prompt with real facts beats a flowery one with none, every time.

Set the boundaries explicitly

Tell it what to do when it doesn't know. My default line is something like: "If you're not certain about a specific detail, say so plainly and suggest contacting the team, rather than guessing." This one instruction prevents the confident-wrong-answer problem that destroys trust. Also tell it what topics to stay away from and when to hand off to a human.

Shape the voice in a few words

You don't need paragraphs about tone. A short, specific instruction does more: "Keep answers short and warm. Use plain language. One or two sentences unless more is needed." I go deeper on brand voice in this piece, but restraint is the theme — over-specifying tone makes it sound like it's trying too hard.

Test it against real questions

Write down the ten questions your customers actually ask and run them through. You'll spot the gaps fast — a missing fact here, a too-long answer there. Tweak, re-run, repeat. Twenty minutes of this is worth more than any amount of theorising.

One more thing: keep the prompt on the server, not in the page. If it's in your front-end code, anyone can read your business context and, worse, try to rewrite the assistant's behaviour. Keeping it server-side is part of the same discipline as protecting your key.

A good prompt is short, specific, factual, and honest about its limits. That's the whole recipe. Everything else is polish.