AI CONTRACT SPEC TEMPLATE (copy/paste into your PRD)

1) Feature name + user job
- Feature: ______________________________
- User job (one sentence): _______________
- Non-goal (what this is NOT): ___________

2) Contract: the user-facing promise
Write this as UI copy you’d be willing to show.
- The assistant WILL: __________________________
- The assistant WILL NOT: ______________________
- It may be wrong when: ________________________
- If unsure, it will: (ask / cite / escalate / stop) _________

3) Inputs + grounding rules
- Allowed data sources (explicit systems): ______________
- Disallowed sources: _________________________________
- Retrieval required? (yes/no; for which intents): _______
- Citation rule: (links, doc titles, inline quotes): ______

4) Actions + tool boundaries
For every tool/action the model can trigger:
- Tool name: __________________
- What it can change: __________
- Scope control: (role-based, allowlist, environment) _____
- Confirmation UX: (preview + confirm? two-person approval?) ______
- Hard blocks (never execute): __________________________

5) Data retention + memory
- What is stored: (prompts, outputs, tool calls, retrieved docs) ______
- Retention period: ____________ (or “configurable by admin”)
- Memory behavior:
 - Session-only mode: (yes/no) ______
 - Persistent memory: (yes/no) ______
 - User controls: (toggle, ‘forget this’, export) __________

6) Safety + refusal behavior
- Refusal categories relevant to this feature: ______________________
- Required safe alternative: (template, FAQ link, handoff) ___________
- Escalation path: (human review queue / ticket / chat handoff) ______

7) Observability (must-have logs)
- Log events:
 - retrieval_performed (doc IDs + ranking)
 - tool_invoked (tool + params + user)
 - confirmation_shown / confirmation_approved
 - refusal_triggered (category)
 - escalation_triggered (destination)
- Redaction rules (what never lands in logs): _______________________

8) Release tests (contract tests, not vibes)
Define invariants; avoid exact-string “goldens.”
- Test case 1: ____________________
 - Invariants: (must cite / must ask / must not call tool / must refuse)
- Test case 2: ____________________
- Regression trigger: run on prompt change / model change / retrieval change / tool schema change.

9) Fallback plan
- If model/API is down: ______________________________
- If retrieval fails: _________________________________
- If tool execution fails: _____________________________

10) Ownership
- Product owner: ___________
- Eng owner: ______________
- Security/Privacy reviewer: _________
- Support runbook link: _____________

Ship rule: if you can’t fill this out without hand-waving, the feature isn’t ready.