What makes a prompt bad.

Most complaints about AI quality trace back to prompt problems, not model problems. Here are the five most common patterns that produce bad replies from a good model.

1. Missing the audience

Bad: "Explain quantum computing."

Better: "Explain quantum computing to a smart CFO who has no physics background but is deciding whether to fund a research initiative."

Reason: The model has to guess. Without the audience specified, the reply is calibrated to the average user's average question — which is the average person, which is nobody in particular.

2. Missing the goal

Bad: "Rewrite this email."

Better: "Rewrite this email to be shorter and more direct — the recipient is busy and I want them to answer yes/no by end of day."

Reason: "Rewrite" is ambiguous. Are you optimizing for tone? Length? Persuasiveness? Clarity? The model will make a choice; better if you make it.

3. Missing the constraint

Bad: "Give me a name for my new startup."

Better: "Give me 5 startup names that are one word, easy to spell over the phone, don't sound like existing tech companies, and could work for a fintech targeting small business."

Reason: Constraints give the model something to optimize against. Wide-open prompts produce wide-open replies that aren't quite anything.

4. Missing the context

Bad: "Review this code."

Better: "Review this code. Context: it's a hot path called 1000x per second in a payment processor. Correctness matters more than elegance. What'd break under load?"

Reason: The model doesn't know the code's role. Review that would be perfect for a hobby project is inappropriate for production. Say what you're actually looking for.

5. Missing the format

Bad: "Compare these three options."

Better: "Compare these three options. Return as a markdown table with columns: name, upfront cost, ongoing cost, ease of migration, notes. Sort by ease of migration ascending."

Reason: The output format is often more important than the content. If you have to reformat every reply, you're wasting time; if you specify format upfront, you get useful output the first time.

Two anti-patterns that look sophisticated

**Overloading a single prompt with 15 rules.** The model attends to some, ignores others, and quality is inconsistent. Better: split into two or three prompts with 3-5 rules each.

**Elaborate role-play framing ("You are an expert X with 20 years of experience...").** Doesn't work reliably. The model isn't imitating an expert; it's producing what a competent language model would produce. Skip the theater; state what you want.

The compound tip

The best-crafted prompts specify all five: audience, goal, constraints, context, format. They're longer than lazy prompts, they produce dramatically better replies, and they're the difference between "AI is fine, I guess" and "AI is a leverage tool."