Improving the result over a few rounds

The first answer from AI is rarely right on the mark, and that is normal. A new colleague is rarely right on the first try either. What separates people who finish quickly from people who go in circles is how they improve it.

The first step is to diagnose, not to repeat. Before you write anything, ask: which part of the checklist made this answer miss? The symptom usually points to one part. If the content is right but the focus is wrong, the Goal is unclear. If the language is too technical or too shallow, the Reader was not named. If there are facts you do not recognise, Material is missing or the model was not told to add nothing from outside. If it is too long or the tone is off, the Limits are not firm. If you have to rearrange the result, the Shape was not named. If you do not know which part to check, Check was skipped.

The second step is specific feedback. "Not good, try again" almost always produces another version with the same problem, because the model does not know what was wrong. Useful feedback has three parts: what is wrong, what to keep and what to change. For example: "Too long and too formal. Keep the three steps and the start date. Make it at most 120 words with the greeting 'Hi team'." One specific round is worth more than five rounds of "try again". Once a fix works, fold it into your original prompt, so next time you do not have to repeat the same round.

The third step applies to prompts you will use again and again, for example to summarise a weekly report or to group customer questions. Before you rely on one, test it on 5 to 10 real examples, including a few hard ones. A prompt that works on one easy example can fail on a messy one. If something misses, fix the prompt, then run all the examples again, not only the one that failed.

If the conversation is already very long and new rounds make the result worse, start a new conversation with the improved request. The reason is covered in the lesson Tokens and the context window.

We call that collection of 5 to 10 examples a test set. Running it and judging the results is called evaluation. It sounds technical, but for office work it is simple.

Choose the examples on purpose. Take two or three of the most ordinary ones, then add hard ones: a very long one, an almost empty one, one in mixed language, one that holds two requests at once, and one that went wrong before. Mask the personal data. Keep them all in one place so you can use them again.

Write down what a "good result" means before you run the prompt, in three to five yes-or-no statements. For example: every figure comes from the material, no more than 150 words, every complaint lands in one group, gaps are written as "not mentioned". If you only think about the measure after seeing the result, you tend to forgive a result that looks tidy.

Change one thing per round. If you change goal, limits and shape at once, you will not know which change helped. Note each prompt version and its results in a simple document.

Remember too that the same prompt can give different answers each time it runs, as explained in the lesson Answers that change. For important examples, run it twice. If the results differ a lot, the prompt is not clear enough yet.