Read a sample: 1. Start with the Symptom
You ask ChatGPT for help with an email. It returns a message that sounds like it was written for a large corporation, even though you run a small neighborhood shop. You ask for a study guide, and it gives you a generic overview instead of covering the material on your test. You ask for a comparison, and the answer confidently recommends an option without considering the one factor that matters most to you.
None of these results necessarily means you have failed at prompting. It means you have something useful to inspect: a response that did not meet your needs.
The temptation is to start over. You might delete the conversation, rewrite the entire request, and hope for a better result. Sometimes that works. But if you do not know what went wrong, you may carry the same problem into the next attempt—or change so many things that you cannot tell what helped.
A more useful approach is to treat the response as a symptom. A symptom points to a problem, but it is not always the problem itself. A stiff email might result from missing information about the audience. An overly long answer might result from an unclear priority. A confident but incorrect claim might reveal a limitation that clearer instructions alone cannot fix.
This chapter gives you a practical cycle for investigating those possibilities. You do not need special terminology or a perfect formula. You need a way to notice the gap between the answer you received and the answer you can use.
The cycle has four steps:
1. Identify the failure. Describe what is wrong with the response in concrete terms.
2. Inspect the request. Look for an instruction, missing detail, or assumption that could have contributed to that failure.
3. Change one or two elements. Revise the request without rebuilding it from scratch.
4. Evaluate the new response. Check whether the change addressed the original problem and whether the answer is still accurate and usable.
You may go through the cycle more than once. The point is not to reach a perfect answer on the first repair. It is to make each revision purposeful.
Suppose your original request was:
> Write an email announcing our new hours.
The response is polished, but it reads like an internal memo to employees. You wanted a short note to customers.
The symptom is not simply “bad email.” It is “the message addresses the wrong audience.” When you inspect the request, you can see that you never said who would receive it. A focused repair might be:
> Write a short email to customers announcing that the shop will now close at 5 p.m. on Fridays instead of 6 p.m. Use a friendly, straightforward tone.
Now evaluate the result. Does it address customers? Does it state the change correctly? Is the tone suitable for your shop? Before sending it, you would also confirm that the hours in the message match the actual schedule.
Notice what changed. You added the audience and the details needed to make the announcement. You did not need a long speech about how the AI should behave. You gave it information the original request left out.
That distinction matters throughout this book. A repair is not necessarily a longer prompt. It is a more useful instruction.
When an answer misses the mark, your first reaction may be “This is terrible” or “That is not what I meant.” Those reactions are understandable, but they do not yet tell you what to fix.
Try finishing this sentence:
I needed the answer to do one thing, but it did another.
For example:
- “I needed a summary of the article I supplied, but the response added general information that was not in it.”
- “I needed a beginner’s explanation, but the response used unfamiliar terms without defining them.”
- “I needed three practical options, but the response gave me a long list with no way to choose.”
- “I needed the message to sound calm, but it sounded accusatory.”
- “I needed verified information about a current requirement, but the response gave me an unsupported answer.”
These descriptions identify different failures. The first may call for a clearer boundary around source material. The second may call for more information about the reader. The third may need a limit and selection criteria. The fourth may need observable tone guidance. The fifth may require checking a reliable source rather than revising the prompt alone.
A useful diagnosis does not have to be elaborate. “Wrong audience,” “too broad,” “missing the main point,” and “cannot verify the claim” are all good starting points. If several things are wrong, choose the most consequential one first. An attractive format is not much help if the answer addresses the wrong question.
Also distinguish between something you dislike and something that prevents you from using the answer. A paragraph may be longer than you expected but still serve your purpose. A confident recommendation may be pleasantly concise but omit a critical uncertainty. Repair should follow your actual need, not just your initial reaction.
Once you have named the symptom, look back at what you asked. Imagine that a capable person received only those words, without access to your intentions. What would they know? What would they have to guess?
Consider this request:
> Help me prepare for my meeting.
It is a reasonable first thought, but it…