Path: AI trainer essentials
Read the instruction and rubric
Turn a broad request into observable checks before judging an answer.
- Starting level
- New learner
- Time
- 18 minutes
- Public lesson guide
- EN + हिंदी
You will learn to
Turn a broad instruction into checks that two reviewers can apply consistently.
Why this matters
Human-feedback work is not about personal taste. A clear rubric turns a vague request into observable evidence.
Three things to remember
- Constraint: what the answer must or must not do.
- Criterion: the quality dimension you will judge.
- Anchor: what pass, partial, and fail look like in the answer.
See one example
Explain a solar eclipse to a 10-year-old in four short sentences.
Make it clear and interesting.
Exactly four short sentences; names the Sun, Moon, and Earth; explains that the Moon blocks sunlight; avoids advanced jargon.
The strong version can be checked in the answer. The weak version depends on taste.
Make something yourself
Write four observable criteria for this task: 'Give a safe, low-cost plan for a student to study during a two-hour power cut.'
Read further
The source is optional. The complete teaching and practice are available inside AI2Bharat.
Read the primary or official source: OpenAI InstructGPT