Most prompts that produce mediocre AI output fail in the same way. They're missing structure.
A good prompt has four parts, in order: role, task, constraints, and output format. Miss any one and the output drifts toward generic. Get all four right and the first response is usually the one you can ship.
1. Role first
Start by telling the model who you are and what you're working on. "I am a product manager working on a checkout flow for a B2B SaaS." This isn't filler. It anchors the model's tone, vocabulary, and what level of detail to assume. Without a role, the model defaults to "helpful generalist," which is usually too cautious and too explainer-y for actual professional work.
The role line is the cheapest upgrade you can make to any prompt. One sentence at the top changes the entire shape of the output.
2. The task, narrowly scoped
"Help me prioritize my backlog" is a bad task. "Score these eight backlog items using RICE with a confidence flag for each score" is a good task.
Specificity isn't picky. It's how you get back something usable instead of something theoretical. If your prompt would produce a different output for two different jobs, narrow it until that's no longer true.
3. Constraints
Length ("under 200 words"). Audience ("for an engineering audience, not for executives"). Tone ("direct, no hedging language"). Without explicit constraints, the model picks defaults that drift toward generic.
Two constraints on every prompt is the rule of thumb. Three is even better. The output gets noticeably sharper.
4. Output format (the part people skip)
"Format as a table with columns Action, Owner, Due date" is dramatically better than "format as a structured response." Modern models follow literal instructions closely. When you leave format ambiguous, they pick the most common interpretation, which is usually flowery prose surrounded by "Here's a summary…" and "Let me know if you'd like more detail!"
Name the sections. Specify the order. Pick the format (table, numbered list, JSON, code block, paragraph). The fewer choices the model has to make for you, the closer the first response gets to what you actually wanted.
The test for any prompt you write
Read your prompt back. Fill in the brackets mentally. Ask yourself: if I sent this to a smart colleague, would they know exactly what to produce?
If the answer is "they'd probably ask me a clarifying question first," the prompt isn't done. Add the missing constraint and try again.
The single biggest upgrade you can make to your prompting today is to stop asking the AI "help me write X" and start asking "I need [specific output], here's the context, here's the format." The first sounds polite. The second produces something you can use.