Prompt engineering sounds technical, but the beginner version is simple: explain the job well. Most weak results come from missing context, unclear success criteria or no requested format.
The CRAFT formula
| Part | Question to answer | Example |
|---|---|---|
| Context | What does the tool need to know? | A customer asked why delivery is delayed. |
| Role | What perspective is useful? | Act as a careful customer-support editor. |
| Action | What exactly should it do? | Rewrite my draft without changing facts. |
| Format | What should the output look like? | Subject + email under 100 words. |
| Tests | How can we judge it? | Acknowledge concern; no invented refund promise. |
Use examples when style matters
If you want a specific format, one short example often works better than three paragraphs of description. Label it as a pattern, not as content to copy. For repeated work, show one acceptable input-output pair.
Ask the model to expose uncertainty
AI tools may fill gaps unless you tell them not to. Add: “If a required fact is missing, write [NEEDS INPUT]” or “Separate facts from assumptions.” This does not guarantee accuracy, but it makes review easier.
Improve prompts one variable at a time
- Run the simplest complete prompt.
- Identify the biggest mismatch.
- Add one constraint or example that fixes it.
- Run again and compare.
- Keep the shorter prompt if both work equally well.
When not to prompt harder
If the source is unavailable, outdated or private, a better prompt does not solve the underlying problem. Get reliable source material first. For high-stakes health, legal, financial or employment decisions, use AI to organise questions—not replace a qualified professional.