Overview
Produces the AI-feature complement to the generic PRD. Every AI-powered FR carries seven mandatory clauses that the generic PRD does not collect. The acceptance gates point at the eval harness as the test oracle.
Quick Reference
| Attribute | Value |
|---|---|
| Inputs | AI_Economic_Value_Brief.md, AI_Feature_Strategy_Doc.md, PRD.md, Multi_Tenancy_Architecture_Spec.md, pricing & packaging spec |
| Output | AI_Feature_PRD_Spec.md |
| Standard | IEEE 830-1998, NIST AI RMF MAP/MEASURE |
Core Instructions
Step 1: Inventory AI-powered functional requirements
List every FR whose output is produced or modified by an AI component (LLM call, RAG, classifier, embedding search, agent action, fine-tune).
Step 2: Attach the seven AI clauses to each FR
For each AI-powered FR, the spec MUST capture:
| Clause | Form | Example |
|---|---|---|
| Hallucination tolerance | factuality score threshold | factuality >= 0.92 on golden set; abstain otherwise |
| Latency budget | P95 target ms | P95 <= 2000 ms; timeout at 8000 ms with graceful fallback |
| $/call ceiling | USD or token cap | <= $0.04/call; throttle when tenant > $X/day |
| Abstain criteria | rule | abstain when retrieval returns < 2 relevant chunks OR confidence < 0.6 |
| Citation policy | rule | every claim about ingested document cites the source span |
| Consent / opt-in | rule | feature is opt-in per workspace admin; default off for EEA tenants |
| Training-data exclusion | rule | tenant content is not used to train the provider model; provider's no-training endpoint used |
Step 3: Define structured output requirements
Where feasible the output is structured (JSON schema, function-call payload, enum). Free-form prose is reserved for user-visible text that has a separate guard (length, style, banned-terms list).
Step 4: Define safety and content rules
State which content policy applies (no medical advice, no legal advice, no investment advice, no PII generation, no protected-class judgements). Cite the safety harness scenarios that verify each rule.
Step 5: Define human-in-the-loop / contestability
Which decisions require human approval before commitment. How a user contests an output (button, reroute, escalation). Reference EU AI Act Art. 14 obligations where the feature is high-risk.
Step 6: Define rollout posture per FR
Initial rollout (canary cohort, opt-in beta, free-tier first), promotion gates (eval pass, red-team pass, hallucination SLO met for N days), and the rollback trigger.
Step 7: Define acceptance tests against the eval harness
Every AI FR has a row in the eval harness golden set with a pass threshold. Acceptance = "harness green for 30 d on this case set + red-team pass for the corresponding adversarial set".
Step 8: Write the spec
AI_Feature_PRD_Spec.md sections: 1) AI FR Inventory, 2) Per-FR AI Clauses, 3) Structured Output Requirements, 4) Safety & Content Rules, 5) Human-in-the-Loop & Contestability, 6) Rollout Posture, 7) Eval Acceptance Gates, 8) Traceability to PRD and to eval harness IDs.
Standards
- IEEE 830-1998
- NIST AI RMF MAP / MEASURE
- EU AI Act Art. 13 (transparency), Art. 14 (human oversight)
- OWASP LLM Top 10
Resources
logic.prompt,README.md,references/ai-feature-prd-spec-template.md,references/ai-feature-prd-addendum.md.