Learn / AI Training Foundations / Module 2 · 11 min

Reading Prompts, Rubrics & Constraints

Good evaluation begins before reading the answer

Before judging a response, understand what the user and the rubric actually require. Many mistakes happen because evaluators remember the topic but miss a constraint.

Constraint decomposition method

  • What is the main task?
  • How many outputs/items are required?
  • What format is required?
  • Who is the audience?
  • What content must be included?
  • What content must be excluded?
  • Are there length, tone, source, date, or language requirements?
Example: convert the prompt into a checklist

Prompt

Give exactly three benefits of exercise. Each benefit must be six words or fewer. Do not mention weight loss.

Checklist: exactly 3 benefits; each ≤6 words; no weight-loss reference.

Prompt

Give exactly three benefits of exercise. Each benefit must be six words or fewer. Do not mention weight loss.

Choose the stronger response.

When instructions conflict

If a prompt asks for “exactly three sentences” and also “one bullet per idea” but requests five ideas, do not invent a solution silently. In real evaluation work, follow the task-specific policy for ambiguity; when allowed, flag the conflict or choose the interpretation supported by the rubric.

Module summary

  • Turn prompts into checklists.
  • Prioritize explicit requirements over personal taste.
  • Separate hard constraints from preferences.
  • Treat ambiguity consistently rather than guessing differently each time.

Quick check

A prompt says “exactly two examples.” The answer gives three excellent examples. Is there an instruction-following issue?

Quick check

What should you do first with a complex prompt?

Quick check

Which is usually more important?