Prompting & Using AI Responsibly

A prompt is just your instruction to the model, but small changes in it produce enormous changes in what comes back. This is the highest-return skill in the whole level, because when an answer disappoints, the fastest fix is nearly always a better prompt rather than a better tool.

The reliable pattern has four parts: role, context, task, format. Role sets the perspective. Context is what the model would need to know and can't guess. Task is the specific thing to do. Format is the shape you want back.

Watch the difference on a real example. The weak version: "write something about our new pricing." What comes back is generic, because nothing in that sentence tells the model who is reading, what changed, or what it's for. The strong version: "You're a product marketer writing to existing customers. We're raising the Pro plan from £15 to £19 a month from 1 March. Existing customers keep the old price for twelve months. Write the announcement email. Lead with the good news about grandfathering, be direct about the increase, no corporate padding. Under 150 words, plain text." Same model, same five seconds. The second one is usable.

Notice what did the work there. Not clever phrasing — information. The specific numbers, the date, the fact about grandfathering, the audience, the length. Most weak prompts are weak because the person knew something relevant and didn't say it. Before you blame the model, reread your prompt and ask what a competent freelancer would have had to email you back to ask.

Four techniques worth having beyond the basic pattern. Show an example: paste one or two samples of the output you want and say "match this style" — for anything with a house format, this beats describing the format in words. Ask for reasoning on hard problems: "work through this step by step before giving your answer" measurably improves multi-step logic, because the model's reasoning happens in the text it writes, so giving it room to work helps. Ask for options: "give me three versions, ranked, with a line on the trade-off" is almost always better than asking for one answer. And set constraints explicitly — length, reading level, what to avoid, what to assume the reader already knows.

Then iterate, because the first answer is a starting position. Don't rewrite the prompt from scratch when something's off; say what's wrong. "Too formal, make it sound like a person." "You invented a discount we don't offer — remove it." "Keep the second paragraph, redo the rest." Precise correction is faster than starting over, and the model has the whole conversation in context.

One habit worth building deliberately: ask it to critique its own output. "What's the weakest part of this?" or "What would a sceptical customer object to here?" Models are noticeably better at spotting problems than at avoiding them in the first place, and this costs one extra message.

Now the responsibility half, which is short and non-negotiable.

Privacy first. Never paste secrets, credentials, or other people's personal data into a tool you don't control. Assume anything you type could be reviewed by a human or retained — check the settings, because most consumer products let you turn off training on your conversations, and business tiers usually exclude it by default. If you're handling customer data, medical records or anything regulated, that decision belongs to whoever owns the compliance risk, not to you individually.

Verification second. Treat output as a draft to check, not a source of truth. Apply the effort in proportion to the cost of being wrong: a brainstormed list needs no checking, a claim you're about to put in front of a customer needs all of it. Be especially sceptical of anything with the shape of a fact — statistics, quotes, citations, case numbers, dates. A fabricated citation looks exactly like a real one. If a model gives you a source, open it.

Honesty third. Be straight about AI's involvement where it matters — where someone is judging your work as yours, where they're relying on human judgement, or where a rule requires disclosure. Using AI to draft is completely ordinary and you don't need to caveat every email. Using it to fake expertise you don't have, or to produce something a person believes a human made when that matters to them, is where people get badly burned.

Do those two things — prompt deliberately, use AI responsibly — and you've cleared the bar for AI Explorer. Prompting has its own full module in Your AI Journey when you want to go deeper.