The shop gives it a voice: AI Engineering, story 2
The same question gets a useless answer, then a great one. The model did not change. See what did.
The model reads a prompt the shop wrote: instructions, examples and room to think shape every answer.
The story, step by step
HM Nayem asks about a jacket
HM Nayem asks the shop's chat about a jacket it sells. In the last story the server passed the question straight to the model. Watch what comes back when it does only that.
A stranger's answer
The reply is polite and useless. The model has never heard of this shop or its jackets. All it saw was one sentence from a stranger.
What the model actually saw
Here is everything the model was sent. One sentence. No shop name, no product list, no hint of who is asking or why. It gave the best answer anyone could give from that.
The shop writes instructions
So the shop writes instructions, and its server places them in front of every question. This is called the system prompt. HM Nayem never sees it, but the model reads it first, every time.
What good instructions cover
Good instructions answer four questions for the model. Who are you? What do you know? How should you sound? What must you never do? Anything left out, the model fills in with a guess.
Same question, new answer
Same question, same model, a completely different answer. Nothing about the model changed. Only the text it was given.
Show it what good looks like
The shop wants shorter replies in one fixed style. Describing a style is hard. Showing it is easy. So the server adds a worked example, and the model copies the pattern.
It copies the pattern
The reply now has the same shape as the example: a direct answer, then one fact. Examples steer a model more reliably than any description.
A question that needs working out
Now HM Nayem asks something that needs a little arithmetic. Pushed to answer in its short style straight away, the model blurts out a guess. The guess is wrong.
Why it guessed
A model writes one token after another and cannot go back. In its short style, the very first word has to be Yes or No. So it must commit before doing any sums. That is close to a coin toss.
Room to think
So the server lets the model think before it answers. It works the sum out in private, then writes the reply. HM Nayem sees only the last part. Thinking costs a little time and a few extra tokens.
The prompt HM Nayem never sees
This is what really reaches the model with every message. HM Nayem wrote one line of it. The shop wrote the rest. Shaping this text is called prompt engineering.
The model answers the text it is given.
- On its own, the model knows nothing about the shop.
- The shop's instructions are placed ahead of every question.
- A worked example fixes the style better than describing it.
- Letting the model think first makes hard answers right.
Ideas in this story
- The system prompt: Instructions the product's builders write, placed ahead of every user message.
- Examples in the prompt: A few worked question and answer pairs. The model copies their pattern.
- Thinking before answering: Letting the model work a problem through before it writes the reply.