AI for Mere Mortals Part 3: How to Write Better Prompts

Large language models (LLMs, the technology behind ChatGPT and similar tools) are trained on huge amounts of text. When you ask a vague question, the model averages across that training data and gives you the most statistically likely answer. The most likely answer to "how do I write a CV" is the boring, generic CV advice you've seen on a thousand career sites. There's nothing wrong with the model. The model just doesn't know anything about you.

The fix is to push the model away from the average. You do that by giving it information it can't have guessed: your specific situation, the audience you're writing for, the constraints you're working within. The more specific the prompt, the more specific the answer.

Every prompt that produces useful output has four things, even if it doesn't look structured. Think of these as ingredients rather than a template: include all four when the result matters, drop the ones that don't apply when you're asking something simple.

The first answer is rarely the final one. Treat the first response as a draft, then push back. The model is very good at refining when you tell it what's wrong. Specific patterns that work:

The model has the full context of your conversation, so each follow-up builds on what's already there. This is the single biggest behavioural difference between people who get useful answers from AI and people who don't: heavy users iterate. They don't accept the first answer and complain about it; they ask for a second pass with specific feedback.

Long conversations accumulate context, which sounds useful but sometimes hurts. After 20 or 30 back-and-forth turns the model starts to lose the thread, repeats itself, or confuses one part of the conversation with another. This is sometimes called "context rot" in the AI community and it's a real effect, not a bug you can fix.

Rules of thumb on when to start fresh: when the task changes completely (you were drafting a CV, now you want help planning a holiday), when the AI seems to be looping on the same wrong answer despite corrections, or when you've made significant changes to your underlying material and want a clean read. Copy the relevant context into the new chat - that's the moment to lean on ingredient one (context) again.

The single biggest unlock for most casual users is realising you can paste large amounts of your own material straight into the chat. Email drafts, meeting notes, your draft blog post, a recipe you're tweaking, a contract you don't understand: it all goes in. The model treats it as part of your prompt.

Patterns that work well:

• Paste then ask. Drop in the email, then say "Rewrite this to sound less defensive and shorter". The model has the full text - it's not guessing what you mean.
• Paste, ask, refine. Drop in your draft, ask for a critique, then ask it to apply the critique. Often the critique pass reveals things you'd never have spotted.
• Paste two things, compare. Drop in two versions and ask which is stronger and why. The reasoning is often more useful than the verdict.

There are obvious privacy implications to be aware of (covered in Part 2) - don't paste sensitive client data or anything legally privileged into the free tier, where conversations may be used to improve models unless you turn that off in settings.