Customers keep bringing in broken prompts — vague, lazy, one-liners — and wondering why the AI gives them junk. You're the mechanic. Pick the right repairs from the toolbox and watch the output transform in front of you. Wrong tools cost strokes. Par is 3 per job. Can you finish all three jobs at par 9?
A prompt isn't magic words — it's a brief. The AI knows nothing about what you want except what you tell it. Context, specifics, and format beat "be creative" every time. Module 04 teaches the full craft.
Most disappointing AI output is not a model failure. It is a briefing failure. The model knows nothing about your situation, your audience, or the shape of the answer you need except what your prompt tells it — and most prompts tell it almost nothing. “Write a poem” is not a request; it is a category.
The good news is that bad prompts fail in predictable ways, and three repairs fix the overwhelming majority of them.
Audience determines vocabulary, tone, length, and structure all at once. A poem for an eight-year-old and a poem for a retirement speech share a form and nothing else. Without an audience the model picks a bland average, which is why unbriefed output so often reads as generically competent and completely unusable.
State the reader, the relationship, and what is at stake. “I am declining a meeting with a long-term senior client I want to keep” produces a fundamentally different email than “write an email saying no.”
This is the repair people skip most, and it costs them the most. Concrete details are what turn a template into a usable draft: the actual reason, the actual constraint, the actual alternative you can offer. An email that says “unfortunately I cannot attend” is a brush-off. An email that says “our product launch lands that same week, and I would rather give your roadmap real attention than half of it — could we do Tuesday or Thursday?” is a relationship being maintained.
The model cannot invent your specifics. Given none, it will produce a shape with blanks in it.
Length, format, and tone constraints are what stop an AI from padding. Asking for “a summary” of 400 customer reviews gets you a paragraph of mush. Asking for “top three complaints, top three praises, and one surprising insight, as bullet points, each with one short supporting quote” gets you something you can take into a meeting.
Requiring evidence has a second benefit that matters enormously: when every claim must carry a quote from the source material, fabricated claims become easy to spot. Structure is not just presentation — it is a verification mechanism.
The game stocks the toolbox with tempting repairs that add zero information, because these are the ones people reach for in real life:
Par is three repairs per job, nine across the shop. Scoring by strokes rewards diagnosis rather than trial and error — the skill is recognising immediately that a prompt is missing its audience, its specifics, or its shape. That instinct is what prompt engineering actually is, and it transfers to every AI tool you will ever use, because it is about communication rather than any particular model’s quirks.
To go deeper, work through Module 04: Talking to AI, which covers few-shot examples, role assignment, chain-of-thought prompting, and iterative refinement.