---
title: "Evaluation-first: why AI products should be tested before they are built"
url: https://manasdasgupta.com/writing/evaluation-first-ai-products/
description: "Why the evaluation set should come before the AI feature, and how to build the first 50 examples."
author: Manas Dasgupta
last_verified: 2026-09-24
---

# Evaluation-first: why AI products should be tested before they are built

> **Bottom line:** Write the evaluation set before you build an AI feature. It forces agreement on what good looks like, turns quality debates into numbers, and catches regressions whenever a prompt or model changes. Start with 50 real examples and add every production failure back in.

> **Placeholder article.** Replace this text with your own. Keep the structure: a BLUF in the frontmatter, then sections that each open with a direct answer.

## What is an evaluation set?

An evaluation set is a list of real inputs paired with the outputs you expect or the qualities a good output must have. It is the AI equivalent of a test suite.

Unlike unit tests, evaluations rarely pass at 100%. The team agrees a threshold, for example 92% of responses graded acceptable, and treats that as the launch bar.

## Why build it before the feature?

Building the evaluation set first forces product, engineering and domain experts to agree what "good" means before anyone is attached to an implementation.

- It exposes disagreements early, while they are cheap.
- It gives the team a number to improve rather than opinions to argue about.
- It becomes a regression gate for every later prompt or model change.

## How to pick the first 50 examples

Pick examples from real usage where possible, and deliberately include the awkward ones.

| Share | Type of example | Why |
|---|---|---|
| 60% | Typical requests | Measures everyday quality |
| 25% | Edge cases | Finds where the feature breaks |
| 15% | Should-refuse cases | Checks the guardrails |

## How to grade

Use the cheapest grader that is reliable for each check:

1. **Exact checks** for format, length and required fields.
2. **Model-based graders** with a written rubric for qualities such as relevance and tone.
3. **Human review** on a sample, to check the graders themselves.

## Key takeaway

If you cannot say how you will measure an AI feature, you are not ready to build it.