How AI systems interpret instructions — and common prompting mistakes

A working mental model

You do not need the mathematics to prompt well, but a simple model of what happens inside helps you predict when a prompt will work — and when to distrust the answer.

1. Your text becomes tokens

The model reads text as tokens: small chunks such as words, parts of words and punctuation. Every model has a context window, the maximum number of tokens it can consider at once (your prompt, the conversation so far and its own reply). Very long conversations or documents can exceed it; then earlier material may be dropped or summarised by the app, and the model can "forget" instructions you gave at the start.

2. Training builds patterns

A large language model is trained on very large collections of text to predict what comes next. Over enormous numbers of examples it learns grammar, style, common facts and common reasoning patterns. Most assistants are then further trained to follow instructions and to respond helpfully and safely.

3. Answers are generated, one token at a time

Given your prompt, the model estimates which tokens are likely to come next, chooses one (usually with a little controlled randomness), appends it and repeats. Three consequences matter for prompting:

• The same prompt can give different answers. Randomness means you should test a prompt more than once before trusting it.
• Plausible is not the same as true. The model can produce a fluent, specific, confident statement that is false — a hallucination. It may invent quotes, statistics, citations or features of a product.
• Everything in the prompt influences the answer. Words you include — examples, tone, even typos and leading questions — steer what is likely to come next. "Why is X the best option?" invites agreement; "Compare X and Y against these criteria" invites comparison.

4. Instructions compete

The model receives instructions from several places: the provider's hidden system instructions, your current message, earlier messages and any pasted documents. When these conflict, results become unpredictable. A pasted web page that says "ignore previous instructions" is text, but the model may treat it as an instruction — this is called prompt injection and you will meet it again in Module 10.

5. The model has a knowledge cutoff

Training data stops at some date. Some tools can search the web or read files you upload, but only when that feature is available and enabled, and retrieved pages can be wrong or out of date.

Common prompting mistakes

Most disappointing outputs trace back to a handful of mistakes:

A first rewrite

Weak: "Write about healthy eating."

Better: "Write a 150-word introduction for a leaflet about healthy eating on a student budget. Audience: first-year university students who cook for themselves. Tone: friendly and practical. Include two cheap meal ideas. Do not give medical advice."

The better prompt names the task, audience, length, tone, content requirements and a limit. You will learn each of these elements properly over the next modules.

Key vocabulary

• Token — a chunk of text the model processes.
• Context window — how much text the model can consider at once.
• Hallucination — fluent output that is false or unsupported.
• Knowledge cutoff — the point after which the model has no training data.
• Prompt injection — text in the input that tries to override your instructions.