Reusable templates, evaluation rubrics and managing uncertainty

From prompt to template

A template is a tested prompt with variables — clearly marked slots that change each time:

Template: Meeting-notes action extractor (v2, tested 3 inputs)
Purpose: turn rough notes into a checked action list.

## Task
Extract action items from the notes below for {{TEAM}}.

## Rules
- Use only information in the notes; if an owner or date is missing, write "not stated".
- Start every task with a verb.
- Do not merge separate tasks.

## Output format
Table: Task | Owner | Due date | Quote from notes

## Notes
<notes>
{{NOTES}}
</notes>

Known limitations: misses tasks phrased as questions; check manually.

Good templates include: a name and version, purpose, variables with a short description, the prompt itself, the expected output format, test notes and known limitations. Store them in a prompt library organised by use case (the library template in your resources shows one way).

Rubrics for evaluating outputs

A rubric turns your success criteria into a scoring guide with levels, so you (or a colleague) can evaluate outputs consistently:

Rubrics are how you compare prompt versions fairly, and they are how your final project — and parts of this course's assessments — are evaluated.

You can ask a model to score an output against a rubric, which can be useful for a first pass. But model-based grading is fallible: it may be lenient, inconsistent or miss factual errors. Spot-check its judgements, and keep a human decision for anything important.

Managing uncertainty and incomplete information

Models tend to answer even when they should not. Build in permission — and instructions — to be uncertain:

• Allow "I don't know" — "If the information is not in the text, write 'not stated' rather than guessing."
• Ask for confidence and gaps — "After the answer, list assumptions you made and information that would change the answer."
• Ask before guessing — "If the request is ambiguous, ask one clarifying question instead of answering."
• Separate fact from inference — "Label each point as 'from the source' or 'my inference'."

These instructions reduce, but do not eliminate, confident errors. Your verification habits from Module 7 still apply.