Prompt Engineering
Prompt engineering is the simple act of giving an AI better instructions. A prompt is the text, image, or request you send in. Better prompts usually lead to clearer structure, better tone, and fewer wasted retries. They do not make the model automatically right.
Where you run into it
You see prompt engineering in everyday AI tools, not just in labs. It shows up when you ask for a draft, a slide, an image, or a quick edit.
- Writing help with Grammarly
- Design help with Canva Magic Studio
- Image generation with Adobe Firefly
- Video editing and generation with Runway
What usually makes a prompt better
Most weak prompts are just too vague. Most better prompts add a few missing pieces. That fits the test-and-review mindset in the Artificial Intelligence Risk Management Framework (AI RMF 1.0) and the human oversight ideas in the OECD AI Principles.
- Say the goal first
- Add the audience or situation
- Ask for a format, like bullets, a table, or steps
- Set limits, such as length, tone, or what to avoid
- Give an example if style matters
- Ask the model to note uncertainty when it is unsure
Try one version, then improve it
Start simple. Then add context, limits, and a clear output shape. The goal is not to write a perfect prompt on the first try. The goal is to compare versions and notice what changed.
A quick test works well: ask for the same task twice, then check which answer is clearer, safer, and easier to use.
Useful does not mean trustworthy
AI is often good at drafting, summarizing, classifying, and rewriting. It is much weaker at checking its own facts. The NIST AI Resource Center is helpful here because it keeps the focus on trustworthy use, not just fast output.
- Usually strong at: first drafts, simple patterns, and format changes
- Usually weak at: fresh facts, rare cases, value judgments, and high-stakes advice
- Common risks: made-up claims, bias, privacy mistakes, and overconfidence
Questions worth asking before you trust the answer
Before you copy, send, or act on an AI response, pause for a minute. That small pause is part of prompt engineering too.
- What is the model being asked to do exactly?
- What key context is still missing?
- Can I verify the main claim with a trusted source?
- What happens if this answer is wrong?
- Did I paste anything private that should not be here?
That last question matters more than people think. The Artificial Intelligence Risk Management Framework (AI RMF 1.0) and the OECD AI Principles both point in the same direction: higher-risk uses need more checking, more records, and more human review.