Iterate, evaluate and improve
The difference between someone good at this and someone frustrated by it is usually a habit of changing one thing at a time.
What you will be able to do
Run a small, disciplined improvement loop on a prompt you use repeatedly.
Before you start
- A prompt you run more than once.
Decide what good means first
Before improving anything, write down what a good output looks like — three to five checkable criteria. Without them you are reacting to vibes, and vibes are not a signal you can act on twice.
One variable at a time
Change one ingredient, run it on the same input, compare. Change three and you learn nothing about which mattered.
For anything you will use repeatedly, keep a handful of test inputs — including the awkward ones. A prompt that works on the easy case and fails on the messy one is not finished.
Let it check its own work
Asking a model to evaluate a draft against explicit criteria catches a useful proportion of problems. It is not a substitute for your review, but it is close to free.
The criteria have to be explicit. "Is this good?" produces agreement, not evaluation.
Step by step
- Write three to five criteria for a good output.
- Collect three test inputs: typical, awkward, and edge case.
- Run the current prompt on all three and note failures against criteria.
- Change one ingredient. Re-run on the same three.
- Keep the version that improves without breaking the others.
- Save the final prompt with its criteria and its test inputs in the same place.
Here is a draft you produced and the criteria it must meet.
Criteria:
1. Under 150 words.
2. No claim that is not in the source I supplied.
3. British English.
4. Ends with one specific next action.
For each criterion, state pass or fail and quote the text that fails. Then produce a corrected version. Do not tell me it is good.What good looks like
- Your criteria are checkable by someone other than you.
- You keep the test inputs alongside the prompt.
- You can say which change made the difference.
What it cannot do
- Self-evaluation shares the blind spots of the generation.
- A prompt tuned on three inputs may not generalise to thirty.
When it goes wrong
- Endless fiddling with no improvement.
- You have no criteria. Stop and write them.
- Improved on one input, worse on another.
- That is why the awkward case is in the set. Optimise across all three.
Privacy and data handling
Test inputs are real data. Keep the awkward ones anonymised if they contain anything sensitive.
Sources2 referenced
- Prompt engineering overview
Anthropic · official guidance · link checked 2026-09-06
- Prompt engineering guide
OpenAI · official guidance · link checked 2026-09-06
Links are checked at build time. If one has moved, tell us.
Read next
The anatomy of a useful prompt
Six ingredients, used as needed. This is a way of thinking about specification, not an incantation you must recite.
Checked 2026-09-06
Structured outputs and reusable prompt patterns
When output feeds a spreadsheet, a script or a colleague, its shape stops being cosmetic.
Checked 2026-09-06
Turn a repeated task into a reusable prompt and context pack
A packaged prompt with its context, examples and evaluation criteria, ready to hand to a colleague.