AI Feature Spec Writer

This skill turns a rough feature idea into a structured AI feature spec using natural language requirements — written in terms of observable system behaviour, not implementation detail.

---

Context

AI feature specs fail for one of three reasons:

Requirements are written for a human engineer, not an AI system

Outputs are described as features ("summarise tickets") not behaviours ("given X, return Y with Z constraints")

Edge cases and failure modes are treated as engineering concerns, not product decisions

---

Step 1 — Extract the raw intent

Ask: what is the feature, who uses it, what triggers it, what should it produce, what must it never do, how will you know it worked.

Step 2 — Write the core requirement set

GIVEN/WHEN/THEN/WITH CONSTRAINTS format. Rules: every requirement must be falsifiable, describe behaviour not implementation, no vague words ("intelligently", "appropriately"), use confidence thresholds instead of "sometimes", one requirement per capability.

Step 3 — Define inputs and outputs explicitly

Structured I/O table: input type/source/format/edge cases, output type/format/destination/latency expectation.

Step 4 — Write failure mode definitions

For each requirement: hallucination, omission, tone violation, over-refusal, latency failure — with specific handling for each.

Step 5 — Write the full spec document

Template: Problem statement, Feature description, Users and context, Functional requirements (GIVEN/WHEN/THEN), Non-functional requirements (latency, accuracy threshold, availability, privacy), Input/Output specification, Failure modes and handling, Acceptance criteria, Out of scope, Open questions.

Quality check before delivering

Suggested next step: Run the aipm-prd-stress-test skill on this spec to find gaps before sharing with engineering.