Vague in, vague out

If your prompt doesn't specify format, length, or behavior on missing data, the model will fill those gaps however it sees fit. And "however it sees fit" rarely matches what you actually need.

The four doors a vague prompt leaves open

1. Format. One sentence? A paragraph? JSON? A list?
2. Length. 20 words? 2 paragraphs? No limit?
3. Language. English? Spanish? The input's language?
4. Behavior on gaps. Invent? Use null? Skip the field?

If your prompt doesn't answer all four, the model decides for you. And the model's decisions are statistically average: you'll get a paragraph in English, variable length, with the gaps filled by whatever sounds plausible.

The fix: read like a skeptic

Before you send your prompt, read it as if you were someone looking for reasons to do the opposite. Every decision that ISN'T explicit is a decision the model takes without you.

Compare:

"Summarize this report."

What language? What length? What do I do if something's missing?

vs.

"Summarize this report in one sentence in English, max 25 words. If info is missing, write 'not reported' instead of inventing."

Four decisions, four explicit answers.

The second prompt isn't longer for the sake of it. It's longer because it closed the four doors the model was about to open.

The catch: don't over-specify

There's a point where more detail stops helping and starts getting in the way. If your prompt has 12 rules, the model will forget some. Practical rule:

• Specify what affects format or correctness of the output.
• Don't specify style, tone, or internal structure unless it matters.

A good heuristic: imagine handing the prompt to a teammate over Slack who has to return the output in 30 seconds. If you'd need to ask 4 clarifying questions first, your prompt is vague. If they'd just get it, you're fine.

On the right, two prompts. One vague, one specific. Which one gives you a more reliable output?