Prompt Engineering Basics: 10 Moves to Make AI 10× More Precise

Most people use AI like this:

You: “Write me a product description.”

AI: “Sure, what’s the product?”

You: “A power bank.”

AI: “Target audience? Selling points? Price tier?”

You (impatient): “Whatever, you decide.”

AI writes generic copy that could apply to any power bank.

You conclude: “AI sucks.”

AI doesn’t suck. Your input was thin. AI’s output is a function of input—how you ask determines how it answers.

These 10 moves each visibly improve quality. All 10 take under 20 minutes to learn.

1. Give it a Role

Letting AI play a specific role instantly elevates output.

❌ “Look at this code, what’s wrong?”

✅ “You’re a 10-year-experienced Python engineer doing code review. Point out the issues in this code: logic bugs, performance, readability.”

Why it works: the model saw vast amounts of “professional content” during training; giving it a role puts it in the matching linguistic space.

2. Give a Goal

Tell it “who this is for, what it’s used for.”

❌ “Write an intro about carbon neutrality.”

✅ “Write an intro about carbon neutrality. For a high schooler with no energy background; goal is to help her understand why it matters in 1 minute.”

Same topic, different goal → completely different output.

3. Give Material

If you have relevant info, paste it directly. Never let AI answer from memory things that can be looked up.

❌ “What were the highlights of our company’s annual report?” (How would it know? It’ll make up.)

✅ “Below is our annual report: [paste]. Extract the three most highlight-worthy items.”

This single move eliminates 80% of hallucination.

4. Give Examples (Few-shot)

For specific style, examples beat 100 instructions.

❌ “Write in a more lively style.”

✅ “Write in this style:

Example 1: ‘Had my morning coffee, instantly transformed from corporate zombie to functional human—caffeine really is civilization’s gift.’ Example 2: ‘Worked till 9pm, came home to find a 6-year-old neighbor had taken 3 bites from my delivery. Not mad, just pondering human nature.’

Now write a paragraph about [topic] in this style.”

This is few-shot prompting—examples beat rules.

5. Step-by-step Thinking (CoT)

For complex problems, don’t ask for the answer—ask it to think first.

❌ “What’s the answer? 25 apples for 4 people, how many each, what about remainder?”

✅ “Reason step by step: 1) divide; 2) compute remainder; 3) discuss how to handle remainder; 4) give final answer.”

Why it works: making AI write out reasoning visibly improves its “thinking” quality—a research-confirmed phenomenon called Chain-of-Thought (CoT).

6. Outline First, Then Write

For long content (article, report, plan), have it outline first, you review the outline, then have it write.

Step 1: Give me a 5-section outline, with each section's content described.
(You review and adjust)
Step 2: Expand the outline into a full article.

Why it works: discovering structure issues at expensive (writing) phase wastes effort. Catch problems at the cheap (outline) phase, then invest at the expensive phase.

7. Give Constraints

Tell it explicitly about length, format, tone, what not to do.

❌ “Write product copy.”

✅ “Write product copy. Requirements:

• Under 100 words
• First sentence must be a rhetorical question
• Start with ‘You’, not ‘We’
• Don’t use words like ‘disrupt’ / ‘redefine’ / ‘ultimate’
• Output 3 versions in: serious / playful / literary”

Why it works: more constraints = more focused output.

8. Let It Ask You

If you don’t fully know what you want—let AI ask you.

✅ “I want to make a report on [X], but haven’t figured out the structure. Play a senior consultant; ask me 5-8 key questions; after I answer, give me an outline.”

Great for “I have a vague idea but can’t articulate.” AI’s questions often surface what you hadn’t thought through.

9. Iterate

Don’t treat AI as a one-shot search engine. Treat it as a colleague who thinks with you.

After first output, at minimum ask:

• “What’s the blind spot in this advice?”
• “If I were an opponent, how would I rebut?”
• “Write a version from a completely opposite angle”
• “Expand on [specific point]”
• “Give me sources for the data/claims here”

Each iteration adds another level of quality.

10. Ask It to Critique

Have AI play harsh reviewer of your (or its) output, in a fresh conversation.

You're a senior editor / boss / mentor known for sharpness.
Below is a draft. With critical eyes, point out:
1. Where's the logic broken?
2. Where's evidence insufficient?
3. Where's language wordy?
4. Where's it unprofessional?
Don't list positives; just problems.
[paste draft]

This move is gold. You often can’t see problems in your own writing—AI playing villain reviewer catches them all.

A Full Example

Combining all 10 moves:

You're a product manager with 8 years of experience (move 1: role).

I just joined a kids' programming education startup, and have to present
to the CEO next month on "product roadmap for the next 6 months"
(move 2: goal + context).

Here's relevant material (move 3: source):
- Current users: 100K parents of elementary kids [more data...]
- Main competitor: [name] has 5x our users
- Our advantage: [list]

Help me do these two steps (moves 5+6: stepwise + outline):

Step 1: List key questions for this kind of planning (move 8: have it ask me),
        I'll answer one by one.
Step 2: After I answer all questions, give me a 5-section outline.

Output constraints (move 7):
- Chinese
- No buzzwords like "empower"/"loop"/"leverage"
- Use specific numbers in bullets

Paste this to ChatGPT/Claude—you’ll find the output quality is completely different.

A Mindset Shift

The real reason most people don’t use AI well: they’re too lazy to type.

Really. They feel “typing this much, might as well write it myself.” But that’s a delusion—

• Typing a 200-word prompt: 2 minutes
• AI generates 2000-word draft: 30 seconds
• You edit: 5 minutes
• Total: 7-8 minutes

If you wrote 2000 words from scratch: at least 30 minutes.

So: “2 minutes upfront prompt investment” vs “22 minutes saved” — fantastic ROI.

Advanced Preview

L4 has dedicated Prompt Engineering content:

• Advanced techniques (self-consistency, tree-of-thought, reflexion)
• Model-specific differences (GPT loves markdown, Claude loves XML tags)
• System prompts
• Prompt injection defense

But master these 10 moves first—they already put you above 80% of AI users.