AI artefact templates and rubrics

Template

Rubric (what "good enough" looks like)

• Scope and boundaries are clear (what is in/out).
• High-impact use is identified (or explicitly excluded).
• Evidence is identified (what you would log/measure/keep).
• Next actions are concrete and realistic.

Worked example (short)

Example (short):
This system drafts customer support replies for low-risk queries.
It must not make account changes, provide legal advice, or handle medical/safety emergencies.
Bad day plan: route to human if confidence is low or retrieved policy is stale; log tool use and citations.

Template

Rubric (what "good enough" looks like)

• Assumptions are explicit and testable.
• Metrics match the real cost of errors (not vanity scores).
• Slices cover a vulnerable group or a high-impact scenario.
• Evidence is identified (what you would log/measure/keep).

Worked example (short)

Example (short):
Primary metric: resolution time saved without increasing complaint rate.
Red flags: citations missing; tool calls spike; user reports unsafe advice.

Template

Rubric (what "good enough" looks like)

• Trade-offs are stated, not implied.
• Monitoring is actionable (alert → owner → runbook).
• Guardrails include permission scoping and rollback.
• Evidence is identified (what you would log/measure/keep).

Worked example (short)

Example (short):
Signals: stale retrieval rate, refusal rate, tool-call anomalies.
Guardrails: approved tools only, per-action limits, fast revoke.