Core responsibilities
- Define instructions, examples and output constraints
- Build representative and adversarial cases
- Document limitations and model-specific behaviour
Technical skills
- Structured instructions and few-shot examples
- Output validation and experimentation
- Analysis of prompt injection and failure modes
Behavioural skills
- Reframe ambiguous requirements
- Prioritise conflicting qualitative feedback
Typical deliverable
Versioned instructions for classifying requests, with an “undetermined” label and a quality rubric.
Possible assessment methods
Improve an instruction on an annotated dataset, then verify side effects on a holdout set.
