# AI-feature extension

A reusable skill the "Go from rough idea to buildable PRD" recipe pulls in. Add it once and every agent on the team can use it.

For AI features, extend the standard PRD with evals, confidence thresholds, and fallback behavior.

## Method
1. Determine whether {{feature_name}} involves a model making a judgment call, generating content, or automating a decision — if so, this extension applies.
2. Add an Evals section: what "correct" output looks like, and how it will be measured before and after ship.
3. Add a Confidence thresholds section: the score or signal below which the feature defers instead of acting (set on first run, then stored in memory).
4. Add a Fallback behavior section: what happens when the model is uncertain, wrong, or unavailable — never a silent failure.
5. Add a Cost-at-scale estimate: token or call cost at expected usage volume, sourced the same way as any other metric.
6. Cross-check that every AI-specific claim (accuracy, latency, cost) has the same sourcing standard as the base PRD's metrics.

## Rules
- Every AI feature gets all four sub-sections — evals, confidence thresholds, fallback behavior, cost-at-scale — none are optional.
- A confidence threshold with no defined fallback behavior is incomplete; both ship together.
- Do not present eval results from a different feature or an aspirational benchmark as if they're this feature's own numbers.
- Cost-at-scale follows the same "real number or [baseline needed]" rule as the base PRD.

## Output
Four added sections — Evals, Confidence thresholds, Fallback behavior, Cost-at-scale — appended to the base PRD structure.