PRD — Output Quality Gate

Use it for

AIA quality gate walkthrough — stop fixing the input, build the gate

The prompt

**Type:** PRD (skill to be built) | **Status:** Extracted — pending design review | **Priority:** High
---
## Problem Statement
AI writing skills produce inconsistent output quality — even well-crafted prompts generate slop on some runs, and model updates reshuffle behavior without warning.
> *"Stop fixing the input. The model is not deterministic. Even great prompts produce slop on some runs... what about if instead of monitoring that, we have a process where we turn it into a number from 0 to 1?"* — Lou
---
## What It Should Do
An external evaluation agent trained on 20–50 gold-standard examples per content type. Runs downstream of any writing skill. Accepts content + a content-type label, scores from 0–1, returns what failed. Never edits. If below threshold, writing skill reruns.
**User journey:**
1. Writing skill produces content
2. Hands to quality gate with content-type label
3. Gate loads matching gold-standard rubric
4. Gate evaluates: score + itemized failure reasons
5. If ≥ threshold: passes. If < threshold: returns failure report → writing skill reruns
6. After N reruns (default 3): passes best-scoring version with flag
7. Gate appends to learning log
---
## Key Design Rules
- Gate scores substance (hook quality, perspective, specificity, angle) — not just grammar/structure
- Gate never rewrites — it scores and blocks
- Multiple rubrics per content type
- Gets harder to fool over time via periodic manual rubric updates
---
## Step Classification
- Load rubric by content-type: code
- Score content against rubric: inference
- Itemize failure reasons: inference
- Compare score to threshold: code
- Write to learning log: code
- Pass/fail routing back to caller: code
---
**Source:** Lou — June 4, 2026 Mastermind (30+ minute demo with live scoring)

Worked example

No worked example recorded yet.