# Day 05: Why vague prompts become vague systems

A vague prompt will still produce a good-looking answer in a demo. That is precisely what makes it dangerous.

The demo input was clean and the ambiguity never got tested. Real users bring the inputs that expose every expectation you left implied, and the model will resolve each one somehow, just not necessarily the way you had in mind.

## What a production prompt is actually carrying

Far more than most teams have written down. Tone. Scope. Output shape. What to do with ambiguous requests. What to refuse. When to say "I do not know." What to do when the evidence is thin or contradictory.

Every one of those is a decision. Left unstated, it does not disappear. It gets made at inference time, differently on different inputs, by something optimising for plausibility rather than your policy.

## Review it like an API contract

The useful discipline is to stop treating prompts as wording and start treating them as operating instructions with acceptance criteria.

A prompt worth shipping states the **objective** (what outcome, judged how), the **constraints** (length, tone, what to never do), the **evidence** it may use, the **output shape**, and the **refusal behaviour**: the last of which is the most commonly missing and the most commonly needed.

Then a specific test: hand the prompt to a colleague and ask what they think the output should look like for a genuinely awkward input. Where their answer differs from yours, you have found ambiguity the model is also resolving on its own.

## Test on the ugly inputs

Prompts get iterated against the cleanest example, because that is what is open in the playground. Then they meet an empty field, a 40-page document, a question in another language, a request that is partly out of scope, or someone being deliberately difficult.

A prompt that behaves only on tidy input is not an operating instruction yet. It is a demo script.

Catch the ambiguity while it is still a wording change. After launch it is a production incident, and by then the vague version is the baseline everyone has adapted to.

## Closing thought

Vague prompts become vague systems. Prompt quality is not about sounding clever. It is about removing avoidable uncertainty before that uncertainty gets resolved for you, at scale, by a probability distribution.

*Day 5 of 60 Days of Production AI Systems.*

---

**Previous:** [Day 04: Why more context can make an AI system worse](https://blog.vamsiannamreddy.com/day-04-why-more-context-can-make-an-ai-system-worse)  
**Next:** [Day 06: Why creativity settings are product decisions](https://blog.vamsiannamreddy.com/day-06-why-creativity-settings-are-product-decisions)  

*Day 5 of 60 · AI Systems Basics (chapter 1 of 10)*