Module 1 of 6 · 20 minutes

How prompts work

A prompt is more than the words you type. Understanding what the model actually receives is the first step to getting what you want.

Learning objectives

  • Describe the parts of a prompt: system prompt, history, context and message
  • Apply the 'new colleague' test
  • Use an iterative prompting loop
  • Explain why different models need different prompting

1.1 What the model actually sees

When you send a message to an AI assistant, the model doesn't just see your words. It receives a block of text (and sometimes images or files) that usually contains:

PartWho writes itWhat it does
System prompt (or developer message)The app builder — or you, through "custom instructions", Projects or custom GPTsSets the role, rules, tone and boundaries for the whole conversation
Conversation historyYou and the modelEarlier turns, which keep influencing later answers
ContextYou, or the app (e.g. via search or retrieval)Documents, data, images and search results
Your messageYouThe current request

The model then predicts a response, token by token, based on all of that. So prompting is really context design: giving the model everything it needs, and nothing that confuses it.

System prompt / custom instructionsset by the app or by you, onceConversation historyearlier turns keep shaping answersContext: documents, search results, imageswhat the model can draw onYour messagethe request right nowModelResponseall of this is 'the prompt'
The model responds to everything in its context, not just your latest message — which is why prompting is really context design.

1.2 The model can't read your mind

Anthropic's prompting guidance suggests imagining the model as a brilliant new colleague with no context on your situation — your organisation, audience, standards or what you tried before. If a smart temp would need more information to do the job well, so does the model.

The colleague test

Before you send a prompt, ask: "If I gave these exact words to a capable colleague who'd just joined today, would they produce what I want?" If not, add what's missing.

1.3 Prompting is iterative

Expert users rarely get the perfect result first time. They work in a loop:

  1. Draft a prompt.
  2. Run it and read the output critically.
  3. Diagnose what's off: missing context? Wrong format? Too long?
  4. Refine — either by follow-up ("make it shorter, and drop the jargon") or by editing the original prompt and running it again.

For one-off tasks, follow-ups are fine. For anything you'll repeat, improve the original prompt so it works first time next time (Module 5).

Draftwrite the promptRunread criticallyDiagnosewhat's missing?Refinefollow-up or editIterate
Expert users rarely get it right first time — they loop. For repeat tasks, fold the fixes back into the original prompt.

1.4 Different models, different strengths

Models differ. OpenAI's guidance notes that reasoning models do well when given a goal and room to work it out — like a senior colleague — while faster general models benefit from more precise, explicit instructions — like a junior colleague. A prompt that works brilliantly in one tool may need adjusting in another, and after a model update. Test, don't assume.

Prompt engineering isn't always the fix

If a task fails because the model lacks information (e.g. your company's latest prices), no wording will fix it — you need to supply the information. And if speed or cost is the problem, a different model may help more than a different prompt.

Key takeaways

  • The model responds to everything in its context — prompting is context design.
  • Write for a brilliant colleague who knows nothing about your situation.
  • Draft, run, diagnose, refine.
  • Reasoning models need goals; faster models need more explicit instructions.