Prompt Engineering Basics

Mental model

A prompt is a contract. You're telling the model:

1. Who you want it to act as.
2. What you want it to do.
3. With what input.
4. In what shape the output should come back.
5. Avoiding what failure modes.

The clearer the contract, the better the output.

Technique 1 — Be specific

Bad: "Write a marketing email."

Better: "Write a 120-word marketing email to existing customers announcing our new annual plan (20% discount, locked-in pricing for 2 years). Tone: friendly but direct. CTA: 'Switch to Annual'. End with one PS line about cancellation policy."

Specificity isn't a stylistic choice. It's the single biggest lever you have.

Technique 2 — Show, don't tell (few-shot)

When you want a specific format or style, give 2–3 examples:

Convert these support tickets into bug reports.

Example 1:
Ticket: "App crashes when I tap settings"
Bug:
  Title: Crash on Settings tap
  Steps: 1. Open app 2. Tap Settings
  Expected: Settings screen loads
  Actual: App crashes

Example 2: ...

Now convert:
Ticket: "{user_input}"

This is more reliable than any amount of instruction.

Technique 3 — Structured output

If you need JSON, ask for JSON and describe the schema:

Return your answer as JSON matching this schema:
{
  "summary": string (max 80 chars),
  "sentiment": "positive" | "neutral" | "negative",
  "action_items": string[]
}
Do not include any text outside the JSON object.

Even better: use the model's native structured output mode (OpenAI response_format, Anthropic tool use, Gemini schema parameter) when available.

Technique 4 — Chain of thought (when it helps)

Asking the model to think step by step before answering improves performance on reasoning tasks. But:

• Modern reasoning models (o3, Claude with extended thinking) do this internally.
• For simple tasks, CoT is wasted tokens.
• Use it when answers are wrong, not as a default.

Technique 5 — Negative examples

Tell the model what not to do:

Do not include speculation. Do not invent statistics. If you don't know, say "Unknown."

These instructions are surprisingly effective at preventing common failure modes.

Technique 6 — Position matters

Models pay strongest attention to the beginning and end of the prompt. Put the most important instructions there. For long context, restate the question at the end:

{long context}

---

Question (repeated for clarity): {actual question}

What doesn't matter as much as you think

• "You are an expert..." preambles — minor effect at best.
• All caps EMPHASIS — sometimes helps, often noise.
• Long elaborate persona descriptions — they pad the context for little gain.

The iteration loop

1. Write the simplest prompt that should work.
2. Run it on 5–10 real inputs.
3. Find the failure mode that bites you most.
4. Add one specific instruction or example targeting it.
5. Repeat.

Prompt engineering is empirical. Trust your eval set over your intuition.