Be Clear and Specific

Turn vague words like "short" and "professional" into instructions a model can follow, and say why you want them so it handles the cases you did not list.

13 minLesson 2 of 8
On this page
  1. Every vague word is a decision you handed over
  2. Watch the guesses disappear
  3. Say why, not only what
  4. Four habits that make prompts specific
  5. Try it yourself
  6. Common mistakes
  7. Check yourself
  8. What to practise next

You ask a colleague to make your slide deck "a bit shorter". They come back proud: 30 slides down to 26. You wanted 10. Nobody was careless. "A bit shorter" had a number in your head, and that number never left it.

Prompts fail the same way. Words like short, simple, professional, a few and improve feel clear to the person typing them. To the reader they are open questions, and a model answers open questions with the most common guess.

Every vague word is a decision you handed over

Prompting Basics showed that a model fills gaps with generic guesses. This lesson is about the gaps hiding inside words you did write. When you say "keep it short", you made a decision (how short) and then did not write it down. The model has to make that decision again, without your reasons.

The fix is to replace each vague word with something you could check. If you cannot tell whether the reply followed the instruction, the model could not either.

VagueSpecific
Keep it shortUnder 80 words
Make it simpleA 12-year-old could follow it; explain any technical word in brackets
Sound professionalFormal, no jokes, no exclamation marks
Give me some ideasGive 5 ideas, one line each
Improve this paragraphFix grammar and spelling only; keep my wording and length

Look at the last row. "Improve" could mean fix typos, restructure, cut by half, or make it sound fancier. Four people would do four different things. "Fix grammar and spelling only" leaves one.

Watch the guesses disappear

Here is a small task: a subject line for an email telling staff the office closes early on Friday. Below are six replies a model might plausibly give to the vague version. All six are fine answers to the question as written. Each detail you add rules some out.

Each detail removes a guessUse the arrows or ← → keys
  1. possible replies
    replydaytimereason
    Office Closing Early
    Important Update
    Early Closure Friday
    Friday: we close at 2pm
    Closing 2pm Fri: power work
    Heads up: early finish
    replies that fit
    6

    1Prompt: "Write a subject line about the office closing early." All six replies answer it. Which one you get is partly chance, as sampling explains.

  2. possible replies
    replydaytimereason
    Office Closing Earlyno
    Important Updateno
    Early Closure Fridayyes
    Friday: we close at 2pmyes
    Closing 2pm Fri: power workyes
    Heads up: early finishno
    replies that fit
    3

    2Add: "Say which day." Three replies never mention Friday, so a reader skimming their inbox learns nothing from them. They are out.

  3. possible replies
    replydaytimereason
    Office Closing Earlyno
    Important Updateno
    Early Closure Fridayyesno
    Friday: we close at 2pmyesyes
    Closing 2pm Fri: power workyesyes
    Heads up: early finishno
    replies that fit
    2

    3Add: "Include the closing time, 2pm." Early could mean noon or 4pm. Only two replies give the time.

  4. possible replies
    replydaytimereason
    Office Closing Earlyno
    Important Updateno
    Early Closure Fridayyesno
    Friday: we close at 2pmyesyesno
    Closing 2pm Fri: power workyesyesyes
    Heads up: early finishno
    replies that fit
    1

    4Add the reason and why it matters: "Mention the electrical work, so people don't think it's a holiday." One reply is left.

  5. possible replies
    replydaytimereason
    Office Closing Earlyno
    Important Updateno
    Early Closure Fridayyesno
    Friday: we close at 2pmyesyesno
    Closing 2pm Fri: power workyesyesyes
    Heads up: early finishno
    details added
    3
    replies that fit
    1

    5Three short additions took the prompt from six acceptable answers to one. The model did not get smarter. You stopped asking it to guess.

1 / 5

In reality there are thousands of possible replies, not six. The principle is the same: each checkable detail shrinks the set of answers that count as correct, and a smaller set means fewer surprises.

Say why, not only what

The last step in the tracer did something different from the others. It gave a reason: "so people don't think it's a holiday". Reasons are worth more than they look, because you can never list every case.

Compare two versions of one rule for a customer support reply:

Rule without a reasonIllustrative example. Real replies vary.
You
Reply to this customer. Never mention refunds. Customer message: My order arrived with a cracked screen. I'd like to return it.
Model
I'm sorry to hear your screen arrived cracked. Please reply with a photo of the damage and your order number, and we'll look into it.

The rule was followed, and the reply dodges the obvious question. The customer asked to return a broken product, and the model had no idea whether that was allowed.

Rule with a reasonIllustrative example. Real replies vary.
You
Reply to this customer. Don't promise refunds: only the billing team can approve them, and promises we can't keep make customers angrier. For damaged items, offer a free replacement, which you are allowed to arrange. Customer message: My order arrived with a cracked screen. I'd like to return it.
Model
I'm sorry your screen arrived cracked. I can send you a free replacement right away. Just reply with a photo of the damage and your order number. If you'd prefer your money back instead, I'll pass your request to our billing team, who handle refunds.

With the reason, the model understood the goal (don't make promises the team cannot keep) and found a helpful answer inside it. It also handled the refund request honestly, a case the bare rule never covered.

Four habits that make prompts specific

  1. Name the reader. "For a new hire in their first week" sets vocabulary, depth and tone in one line.
  2. Give numbers. Word counts, number of items, reading level, a deadline. Numbers are the easiest instruction to follow and to check.
  3. Describe done. What does a good answer contain? "It should end with one action the reader can take today."
  4. Say what to do when stuck. If the input might be missing something, say how to handle it: "If the date is not in the text, write unknown; do not guess." Without this, the model fills the gap with something plausible, which is how made-up details slip in.

Try it yourself

Try it: Make the product description specific

This prompt produces a generic, gushing paragraph. Rewrite it. Replace every vague word with something checkable, name the reader, and give at least one reason.

Show a sample answer

One stronger version:

Rewritten promptIllustrative example. Real replies vary.
You
Write a product description for our online shop. Product: steel water bottle, 750 ml, keeps drinks cold for 24 hours, fits standard car cup holders, dishwasher safe, 3 colours. Reader: people who cycle or commute and are comparing several bottles on their phone. They skim, so put the most useful fact first. Length: 50 to 70 words, then 3 bullet points with the key specs. Tone: plain and confident. No exclamation marks and no words like "revolutionary" or "game-changing", because our customers have told us that kind of language makes them trust a listing less.

Why it works:

  • "Good" and "catchy" became a reader (skimming commuters) and an ordering rule (most useful fact first).
  • "Not too long" became 50 to 70 words plus 3 bullets, which you can check at a glance.
  • The facts are given, so the model does not invent features.
  • The ban on hype words comes with a reason, so the model also avoids hype words you did not list.

Common mistakes

Check yourself

Quick check

Which instruction is the most specific?

Why does giving the reason behind a rule help?

Your prompt says "Extract the invoice date." Some invoices have no date. What should you add?

What to practise next

Some things are hard to describe in words at all, like a tone or a tricky format. For those, showing beats telling: continue with Show Examples.