You open an AI chat, type “help me plan a birthday party,” and get back twelve bullet points that read like a greeting card written by a committee. Nothing is wrong. No decisions were made. You close the tab and go back to doing it yourself.
The tool isn’t the problem. The briefing is. An AI chat has no idea who you are, what you’ve already tried, or what a good answer looks like in your world. It will happily fill that silence with the most average version of whatever you asked for.
So let’s fix that with one ordinary task, start to finish. I’ll use chasing a landlord about a broken boiler, because it’s boring, specific, and the kind of thing most of us have actually done.
Write the brief before you write the prompt
Most people type the request. The people who get good results type the context first, then the request. There’s a reason for that order: context shapes everything that follows, and once the model has committed to a generic answer it tends to stay there.
Four lines cover almost any task:
- Who you are. “I rent a flat in Leeds, tenancy started in March.”
- Who it’s for. “This goes to a letting agent who has ignored two previous emails.”
- What done looks like. “A 120-word email that gets a repair date, not an apology.”
- What’s off-limits. “No threats of legal action yet. No exclamation marks. Don’t mention the deposit.”
That’s maybe 30 seconds of typing. It’s also the single biggest difference between a usable draft and a paragraph you delete. If you want a fuller framework, this step-by-step guide to getting useful answers from an AI chatbot breaks the same idea into eight moves you can apply to almost anything.
The before and after, side by side
Here’s the prompt most people send:
Write an email asking my landlord to fix my boiler.
The reply is polite, vague, and useless. Something about “kindly requesting” and “at your earliest convenience.” It could have been written for anyone on the planet.
Now the briefed version:
I rent a flat in Leeds. The boiler has been dead for nine days, so no hot water. I’ve emailed the letting agent twice with no reply. Write a 120-word email, firm but not aggressive, that asks for a repair date this week and notes that I’ve kept a record of the delays. Plain British English. No exclamation marks.
Same model, same chat window, completely different output. The second draft actually names the nine days, asks for a date rather than a vague commitment, and stops the reply from drifting into corporate blandness. That’s not the AI being clever. It’s you removing its excuses.
Work in rounds, not one big shot
A single prompt gets you a starting point. The value shows up in rounds two and three, when you stop accepting and start steering.
Round one: options
Ask for three versions with different temperatures. “Give me a neutral one, a firm one, and one that assumes they’re about to fix it.” You get range instead of a single guess, and range is what lets you pick.
Round two: critique
This is the move almost nobody makes. Follow up with something like: “Which of those three is most likely to get ignored, and why? Rewrite it so that weakness is gone.” The model is decent at spotting its own filler when you point at it directly.
Round three: reverse the roles
“You’re the letting agent. You’ve got 40 open tickets. What’s the fastest reason you’d push this email to the bottom of the pile?” That answer is often more useful than the draft itself, because it tells you what to cut.
Make it show its working
AI chat is confident in a way that has nothing to do with being right. Ask for a legal deadline and you may get a specific number of days delivered with total calm, whether or not that number applies in your country.
The fix is mechanical. When a factual claim matters, ask two follow-ups: “Where does that come from?” and “What would change your answer?” If it can’t point anywhere checkable, treat the claim as a draft, not a fact. There’s a useful rundown of the common failure modes in this guide to how AI chatbot GPT behaves and how to use it without getting fooled, and it’s worth ten minutes before you rely on a chat for anything with money or deadlines attached.
Numbers deserve their own check. If you ask for a budget split or a repayment schedule, request the arithmetic step by step. Models slip on multiplication far more often than on prose.
Keep the thread warm, but know when to walk away
Long conversations are usually better than fresh ones. The chat already knows your flat is in Leeds and that you dislike exclamation marks, so you don’t have to repeat yourself.
Two exceptions. First, if the answers start getting longer and more agreeable, the thread has drifted. Restart with your four-line brief pasted in. Second, if you’re switching tasks entirely, start clean. Context from a boiler complaint will leak into the birthday invitations and make them sound weirdly formal.
One trick worth stealing: paste in 200 words of your own writing and say “match this voice.” A sample beats any description of tone. “Write in a friendly tone” means nothing; three of your own sentences means everything.
Match the tool to the job
Not every AI chat is equally good at every task. Google’s assistant tends to do better when you need something current or sourced from the live web, and it’s worth understanding what Google’s AI chatbot does well and where it lets you down before you lean on it for research-heavy work.
For long drafting, rewriting, and anything where you’re iterating on tone, a general-purpose chat model usually holds up better across a long thread. If you’re deciding where to spend your time, this no-hype comparison of the main AI chatbots is a sensible starting point rather than a sales pitch.
One more thing: these tools shift under you. Features that were broken last spring work now, and the reverse happens too. Keeping half an eye on what’s improved and what still breaks in ChatGPT saves you from abandoning a workflow for a reason that stopped being true.
The traps worth memorising
You’ll hit these repeatedly, so name them now:
- Invented sources. Real-looking titles, wrong authors, pages that don’t exist. Never cite a chat’s reference without opening it.
- Confident arithmetic. Percentages and totals drift. Ask for the steps.
- Flattery. “Great question” and “excellent point” are filler. Ignore them and keep steering.
- Stale facts. Anything after the training cut-off needs checking against a live source.
- The average answer. If a draft could apply to any business, any flat, any customer, it hasn’t been briefed properly.
Build a prompt library you’ll actually reuse
After a few weeks you’ll notice the same three or four moves working again and again. Write them down in a notes app, not in your head.
The snippets that earn their place, in my experience:
- “Here’s my brief. Ask me three questions before you answer.”
- “Give me three versions: cautious, direct, and confrontational.”
- “List the three weakest claims in this and what evidence would fix each.”
- “Rewrite this at 60% of the length without losing a single fact.”
- “Match the voice of this sample: [paste 200 words].”
That’s the whole method. Brief properly, ask for options, attack your own draft, verify anything with a number or a date in it. The last 10% of quality comes from you reading the output like a sceptical editor rather than a grateful recipient. Do that, and AI chat stops being a novelty you poke at and becomes the fastest first draft you’ve ever had.