Practical AI training for everyone
LearnModule 3
6 min read
Most people's first disappointing AI experience has the same cause. They typed something like "write me a report on climate change," got back four paragraphs of bland Wikipedia-flavoured text, and concluded the tool was overhyped.
The tool did exactly what was asked. "Write me a report on climate change" describes ten thousand possible documents, so it produced the statistical average of all of them, which is, by definition, generic. Generic input, generic output. The fix is not magic words. It is specificity.
A prompt that works almost always contains four things. You do not need to label them or put them in order: you just need them present.
Weak
Write a report on climate change.
Strong
You are a secondary school science teacher. Write a one-page briefing for parents explaining what their 14-year-olds will study in our new climate unit. They are not scientists and some are sceptical of the topic, so stay factual and avoid anything that reads as political. Three short paragraphs, warm and plain, ending with one sentence on how they can help at home.
The second is longer to type and produces something you might actually send. That trade is almost always worth it, and note that nothing in the strong version is a trick. It is just the information a competent human would have needed to do the job.
Before you send a prompt, ask: could a capable new colleague do this well with only what I have written? If they would have to come back and ask you a question, the model will not ask; it will guess, silently, and you will get the consequences of that guess.
This one test catches the majority of weak prompts, because the missing ingredient is nearly always context you have in your head and forgot was not on the page.
Show, do not just tell. If you have an example of what good looks like — a previous email, a paragraph in your voice, a report in the right format — paste it in and say "match this style." One example is worth several sentences of description, because you almost certainly cannot articulate your own writing style accurately, but you can recognise it.
State the negatives. Models drift toward corporate filler by default. "No bullet points. Do not use the words 'delve', 'leverage' or 'robust'. Do not open with a summary of what I just asked" costs you one line and removes the three things you would otherwise edit out by hand every time.
Try it now
Take a real task from your week. Write the lazy one-line version of the prompt, then rewrite it with all four ingredients. Run both. The gap between the two outputs is the actual return on learning this skill, and it is usually much bigger than people expect.
A few quick questions. You'll see an explanation after each one.
Question 1 of 5
Why does 'write a report on climate change' produce bland output?