Skip to content
Tutorials Center
My learning

Topics

Lesson 5 of 38 ยท Prompting Fundamentals

Writing Clear, Specific Instructions

Turn vague requests into specific, testable instructions: say what to do, how much, for whom and what to avoid. Includes a runnable specificity checker.

The single biggest improvement most people can make to their prompts is to be specific. A model cannot read your mind or see your screen. It sees only the words you send. If a requirement matters to you but is not in the prompt, treat it as missing.

Specific does not mean long. It means each sentence removes a decision the model would otherwise make for you.

Say what to do, not only what to avoid

Positive instructions are easier to follow than lists of prohibitions. "Don't be wordy" is vague; "Use at most three sentences" is concrete.

Vague Specific
Make it shorter Cut it to 50 words, keep the refund amount
Be professional Use a polite, neutral tone; no exclamation marks
Explain it simply Explain it to a 14-year-old using one everyday analogy
List some ideas Give exactly five ideas, one line each
Fix the code Find the bug, explain it in two sentences, then show the corrected function

When you do need a restriction, pair it with the alternative: "Do not use slang; write in standard business English."

Make requirements measurable

A good instruction can be checked. If you cannot tell by looking whether the output obeyed, the instruction is too fuzzy.

  • Length: words, sentences, bullets, paragraphs.
  • Count: "exactly 3", "at least 2", "no more than 5".
  • Content: "must mention the deadline", "must not include prices".
  • Form: JSON keys, headings, a table with named columns.
  • Language and tone: "UK English", "formal", "second person".

Measurable instructions also let you write automated checks later, which is the foundation of the evaluation chapter.

Spell out the steps when order matters

If the task has several parts, number them. Models follow explicit sequences more reliably than they infer them.

TEXT
Read the review below, then:
1. Classify the sentiment as positive, negative or mixed.
2. Quote the single sentence that best supports your classification.
3. Write a one-sentence reply to the customer.

Remove hidden assumptions

Read your prompt as a stranger would. Phrases like "the usual format", "as before", "our style" mean nothing to a model unless you supply them. Replace each with the actual content or a short description.

Python Check a prompt for vague words (heuristic stand-in)

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

This is only a smoke test: a prompt can avoid every word on the list and still be unclear. But the habit it models, scanning for fuzzy adjectives and missing numbers, is real.

Tell the model what to do when it cannot

Models tend to guess rather than admit uncertainty. If a field might be absent or a question unanswerable, say what to do:

TEXT
If the text does not state the delivery date, write "unknown".
Do not guess a date.

This one sentence prevents a classic failure: an invented but plausible-looking value.

Show a before and after

A realistic rewrite. The vague version might return a paragraph, a list, or a question. The specific one narrows the space of acceptable answers to a handful.

TEXT
Before:
Summarize this meeting.

After:
Summarize the meeting transcript below for people who did not attend.
Output three sections with these exact headings: "Decisions", "Open questions",
"Action items". Under each, use bullets. Each action item must name an owner and
a date; if either is missing from the transcript, write "unassigned".
Maximum 150 words in total.

Instruction hygiene in code

If instructions live in a template, keep each rule on its own line. That makes diffs readable and lets you toggle rules individually. The block below assembles a rule list and proves a property you care about: no rule is accidentally duplicated.

Python Keep rules as a list, not a paragraph

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Do not confuse specific with rigid

Over-constraining can backfire. Twenty overlapping rules compete and conflict, and the model may satisfy some and drop others. Prioritize: state the two or three requirements that matter most, then test. Add rules only to fix observed failures.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. Which instruction is the most specific?

    A word limit is measurable, so you can verify whether the output obeyed.
  2. 2. What should you tell the model to do when information is missing?

    Without a fallback instruction, models often invent plausible values.
  3. 3. Why number the steps of a multi-part task?

    Explicit sequences reduce skipped or reordered steps.
  4. 4. What is the risk of adding twenty overlapping rules?

    Prioritize the key requirements and add rules only to fix observed failures.

Summary

  • Be specific: every sentence should remove a decision the model would otherwise guess.
  • Prefer positive, measurable instructions (counts, lengths, headings) over vague adjectives.
  • Number the steps when order matters, and say what to do when information is missing.
  • Keep rules as a short, prioritized list and add more only when testing shows a failure.

Next: Giving Context and Naming the Audience