Testing and fixing it

Your prompt is installed. It probably works about eighty percent as well as it could, which is already a huge improvement over nothing. The last twenty percent comes from testing it the way you would test a new employee: give them the same five tasks, look at what came back, and adjust the briefing, not the person.

The key word is adjust. People who get frustrated with AI tend to do one of two things: give up, or throw away the whole prompt and write a new one from scratch. Both waste the work you have done. A system prompt is a living document. You will edit it a dozen times in the first month and then maybe twice a year after that.

The five-question test

Run these five in a fresh chat, one at a time, every time you change the prompt. They are designed so that each one stresses a different part of the briefing. Write your own versions, but keep the shape.

1. The identity check. "Who am I and what do you do for me?" Tests whether the role, audience, and goals landed at all. If the answer is generic, the prompt is not installed or the scope is wrong (see lesson 5).
2. The typical task. The most common thing you will actually ask for. "Write a reply to a five-star review from Karen who said the crew was on time and cleaned up." Tests voice, length, and format rules.
3. The pressure test. Ask the thing a real customer would ask that pushes on a boundary. "A customer wants to know roughly what a full tint job costs. Just give them a ballpark." Tests whether the boundaries hold under a polite push.
4. The missing-detail test. Give a task with a hole in it. "Write a follow-up to the customer from yesterday." Tests whether the model asks instead of inventing a name and a job.
5. The override test. Deliberately ask for something the format rules forbid. "Write this one as a bulleted list." Tests whether the prompt treats format rules as defaults (it should comply) rather than absolutes (it should not lecture you).

Grade each answer as pass, close, or fail. Anything that fails, note which part of the prompt should have caught it. Most people find one or two fails on the first run, usually on questions 3 and 4.

The common failure patterns

Nearly every misbehaving system prompt has one of a small number of problems. Learn to recognize them and you can diagnose a prompt in a minute. Tap the highlighted parts of this one to see each problem named.

Here is the full list, with the fix for each.

• Too long. The prompt rambles, and instructions in the middle get skimmed. Symptom: rules near the top and bottom work, rules in the middle do not. Fix: cut anything that would not change the output if deleted. Move the rules that failed to the end, right before your message.
• Contradictions. Two rules pull in opposite directions (formal and casual, brief and thorough). Symptom: the output is inconsistent from chat to chat, obeying one rule one day and the other the next. Fix: pick one, or say when each applies: "casual on Facebook, formal in email."
• Vague persona. "Be professional and friendly." Symptom: output that is fine and sounds like nobody. Fix: replace adjectives with a real sample, as in lesson 3.
• Rule with no replacement. "Never discuss pricing." Symptom: stiff refusals that end the conversation. Fix: add what to do instead.
• Dangerous positive rule. "Always give an exact number." Symptom: confident invention. Fix: turn it into a boundary and a hand-off to you.
• Facts the model cannot know. The prompt assumes the model knows your hours, your service area, your price list. Symptom: it makes them up. Fix: a short FACTS block in the prompt, or a file attached to the project.

The iteration loop

Here is the routine. It takes about ten minutes per round and you will do it three or four times in the first week, then rarely.

1. Run the five questions in a fresh chat. Fresh matters: a long existing chat has history that muddies the result.
2. Pick the single worst failure. Not all of them. One. Fixing several at once means you cannot tell which change worked.
3. Find the sentence in the prompt that should have caught it. If there is one, sharpen it: more specific, with a replacement behavior. If there is not one, add one sentence, and consider what you can delete to make room.
4. Save the new version in both places: the app, and your own notes. Date it. You will want to see how it evolved.
5. Rerun only the question that failed. If it passes, rerun all five to make sure you did not break something else. If it fails, try a different wording before adding a second rule.

Two more habits help. First, keep a scratch list of real outputs you did not like during normal use, with a note about what was wrong. Those become the next round's tests. Second, once a month, ask the AI itself to review the prompt: "Here is my system prompt. Point out any contradictions, vague instructions, or rules that do not say what to do instead." It is good at this, because it is reading the prompt the way it always does, and it will tell you which parts are hard to follow.

When to stop

You are done, for now, when all five questions pass two runs in a row and you have gone a week of normal use without a fix you felt strongly about. That is not perfection. It is a briefing that works well enough that you have stopped thinking about it, which was the goal from lesson 1.

Two things will bring you back. The first is your business changing: a new service, a new town, a new tone after a rebrand. Update the prompt the same day you update the website. The second is the app changing: models get replaced and a prompt tuned for one may drift on the next. When your tool announces a new model, rerun the five questions. It takes ten minutes and usually nothing needs to change.

That is the whole module. You have gone from a hidden message you did not know existed to a written, installed, tested briefing that makes every chat start on your terms. The next module, Working with Documents and Data, picks up where this leaves off: giving the assistant your actual files, price lists, and records to work from.

Try this yourself

Do your first test round now. Open a fresh chat inside your installed setup and paste this, filling in the brackets with your own versions of the five questions.

I am testing your setup. Answer each of these separately, one after another, without commentary between them.

1. Who am I and what do you do for me?
2. [Your typical task, with a real detail: a name, a job, a review.]
3. [A customer pushing for a price, date, or promise.]
4. [A task with a detail missing, like "follow up with yesterday's customer."]
5. [A request that breaks a format rule on purpose, like "do this one as a bulleted list."]

Grade the five answers. Then fix the worst one, and only that one, using the loop above. Save the new version with today's date. You have just done something most people who use AI every day have never done: you checked whether it was actually listening.