What you'll learn
If you learn only one prompting skill, make it this one: say exactly what you want. Most disappointing answers aren't the model's fault — they're the predictable result of a vague request. This module is a toolkit for writing instructions so clear the model can't misread them.
By the end of this module you'll be able to:
- Turn a vague request into a specific, unambiguous one
- Frame instructions as what to do, not just what to avoid
- Break a tangled ask into ordered, single-purpose steps
- Give the model a safe "out" so it stops inventing answers
Be specific, not vague
Recall from Phase 1 that anything you leave unstated, the model fills in with an average guess. Specificity replaces those guesses with your actual intent. Watch what happens when we pin down the audience, the format, and the goal:
Prompt
Give me feedback on my resume.
Response
Prompt
You are a technical recruiter. Review the resume below for a SENIOR BACKEND ENGINEER role. Give exactly 3 specific improvements, each quoting the line it refers to, then rate ATS-friendliness from 1-10 with one reason.
Response
Key idea
| Vague | Specific |
|---|---|
| “Make it better” | “Shorten to 150 words and use active voice” |
| “A few examples” | “Exactly 3 examples” |
| “Soon” | “By the end of the first paragraph” |
| “Professional tone” | “Formal, third-person, no contractions” |
| “Summarize this” | “Summarize in 5 bullet points a busy manager could skim” |
Say what to do, not only what to avoid
Models follow positive instructions more reliably than negative ones. Telling a model what not to do still leaves infinite room for what it might do instead. Whenever you catch yourself writing "don't," try to restate it as a "do."
| Instead of (negative) | Prefer (positive) |
|---|---|
| “Don't be too long” | “Answer in 2-3 sentences” |
| “Don't use jargon” | “Explain it so a 12-year-old understands” |
| “Don't make things up” | “Only use facts from the text below; otherwise say ‘not stated’” |
| “Don't be informal” | “Use a formal, professional tone” |
One task at a time (or number them)
When a prompt bundles several requests into one run-on sentence, the model often drops one. If a task has multiple parts, number them — it gives the model a checklist to work through and makes it obvious to you if a step was skipped.
Prompt
From the review below, do all three, labeled: 1. Overall sentiment (positive / neutral / negative) 2. The single biggest complaint, in under 10 words 3. A one-sentence reply we could send Review: "Fast shipping, but the app crashes every time I open settings."
AI response
2. Biggest complaint: app crashes when opening settings
3. Reply: Thanks for the kind word on shipping — we're sorry about the settings crash and are pushing a fix this week.
Give the model an out
A model will often rather guess than admit it can't answer — that's a recipe for hallucination. The fix is simple: explicitly permit "I don't know" and tell it what to say instead. This one clause prevents a huge share of confident-but-wrong answers.
Prompt
Using ONLY the bio below, what is the author's nationality? If the bio doesn't say, reply exactly "Not stated" — do not guess. Bio: "Dr. Lena Okafor is a marine biologist who has published over 40 papers on coral reefs."
AI response
Tip
Common instruction mistakes
A quick checklist of the traps that quietly ruin otherwise good prompts:
- Burying the ask. The main instruction is hidden in paragraph three. Put it up front.
- Conflicting instructions. "Be thorough but keep it to one line." Pick one.
- Assumed context. Referring to "the usual format" the model has never seen.
- Undefined terms. "Make it pop" means nothing measurable.
- No success criteria. If you can't tell whether the output is right, neither can the model.
Recap & quick check
Key takeaways
- Specificity is the highest-leverage prompting skill — replace every vague word with something concrete.
- Prefer positive instructions ('do this') over negative ones ('don't do that').
- Split multi-part tasks into numbered steps so nothing gets dropped.
- Give the model an explicit out ('say “not stated” rather than guess') to curb hallucination.
- Put the main instruction up front, avoid conflicting or undefined terms, and state success criteria.
Quick check
1. Why does 'Give me feedback on my resume' tend to produce generic advice?
2. Which instruction will the model follow most reliably?
3. What's the benefit of numbering the sub-tasks in a prompt?
4. How does 'giving the model an out' reduce hallucination?
Clear instructions tell the model what to do. Next we tell it who to be — and see how a role and a system prompt shape everything that follows. Next up: Module 6 — Roles, Personas & System Prompts.