Policy + Prompt Pack (Starter)

- Human-in-the-loop for anything external
- Red/Amber/Green data classification; only Green in consumer models
- Prefer client enterprise tenancy; otherwise use local/offline tools
- Cite sources; run chain-of-verification for facts
- Keep audit trail (prompt, output, reviewer, decision)

Prompt Patterns (copy & adapt)

Role-Task-Context-Constraints-Format
You are a [role].
Task: [what to do].
Context: [facts, source].
Constraints: [brand, policy].
Output format: [bullet spec].
Summarize-Then-Decide
1) Summarize the input faithfully (no advice).
2) Given the summary, decide [X] and justify with quotes.
Extract-Structure
Extract fields into valid JSON: { "field": "type", ... }.
If missing, use null. Return ONLY JSON.
Compare-Against-Rubric
First critique draft vs rubric. Then rewrite to meet rubric.