How Language Models Read Your Prompt

A large language model (LLM) such as Claude or GPT predicts text. Given everything in its context (your instructions, documents, the conversation so far), it writes the most fitting continuation. Prompting well means giving it the context a smart new colleague would need to do the job.

Tokens and context windows

Models read tokens: chunks of words (roughly ¾ of an English word each). Every model has a context window, the maximum tokens it can consider at once, and you pay per token in and out. Current frontier models handle very long contexts (hundreds of thousands of tokens or more), so pasting whole documents is normal; but more irrelevant text still means more cost and more chances to focus on the wrong thing.

Why models "make things up"

A model fills gaps with plausible text. If you ask for a statistic it does not have, it may produce a convincing but wrong one. You reduce this by:

• giving it the source material,
• telling it what to do when information is missing ("If the document does not say, answer 'not stated'"),
• asking it to quote or cite the text it relied on.

The mental model: a brilliant new hire

Imagine handing the task to a capable colleague on their first day. They know a lot in general but nothing about your project, audience, constraints or definition of "done". Everything they need must be in the prompt. Most weak outputs come from missing context, not from the model's ability.

What a good prompt contains

1. Goal: what you want and why it matters.
2. Context: who the audience is, background, source material.
3. Constraints: length, tone, format, things to avoid.
4. Examples (when the format or style is subtle).
5. Output format: exactly how the answer should look.

Before and after

Weak:

Write a product description for our hoodie.

Strong:

Write a product description for an online store.

Product: "Midnight Panda" hoodie, 380 gsm organic cotton, oversized fit,
embroidered logo, sizes XS–3XL, $59.
Audience: developers aged 20–35 who like understated design.
Tone: friendly, confident, no hype words like "revolutionary".
Length: 60–80 words, then 3 short bullet points of key specs.

The second prompt removes guesswork, so the first answer is usually usable.

Check yourself

• What is a token, and why do token counts matter?
• Name two ways to reduce made-up answers.
• What five parts make a prompt complete?