Prompt Engineering Techniques

Before you tweak the prompt

Decide what “good” looks like. Write down success criteria, 3–5 test cases, and a word/length target. Then iterate against that set so improvements are real, not vibes.

The core moves that matter

These three techniques do most of the work:

1. Role + guardrails

• Give a clear job and boundaries: who are you, what to include/exclude, how to format.

2. Tiny examples (few‑shot)

• One or two short examples calibrate tone, structure, and level of detail.

3. Structure first

• Pre-fill the shape (headings, bullets, or JSON keys) so the model completes it, not invents it.

Quick recipes

• Improve adherence: add a role, restate constraints at the end, and pre-fill headers.
• Reduce waffle: ask for bullets and a hard word limit; request a short takeaway.
• Boost reasoning: “Think step-by-step, then give a 2‑sentence final answer.”
• Compare options: “Propose 2 approaches with pros/cons and a recommendation.”
• Extract facts: use tags and ask for structured fields you actually need.

Mini templates

[Role]
You are a careful, practical {persona}. Stay within the provided context.

[Context]
Audience: {who}
Goal: {what}
Constraints: {length, tone, exclusions}

[Instructions]
1) Outline the answer in 3 bullets.
2) Provide the answer.
3) End with a 1‑line takeaway.

[Output]
- Bullets
- Answer
- Takeaway

[Critique + Improve]
Task: {brief}
Current prompt: "{prompt}"
Do: list 3 concrete issues, then propose a revised prompt. Keep ≤120 words.

Try it live

Try expanding the initial system prompt and see how the AI model reacts!

Key Takeaways:

• Role + guardrails, tiny examples, and structure-first do most of the work.
• Define “good” up front and iterate against test cases.
• Pre-fill the shape so the model completes rather than invents.

More Resources:

• Prompts 101
• Understanding System Prompts

Sources:

• Anthropic - Prompt engineering overview: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview