Content Expert Panel & Quality Gate

What it does

The Content Expert Panel is a general-purpose scoring and recursive-improvement engine. For any artifact you hand it (copy, email sequence, landing page, strategy doc, title, chart, candidate eval) it auto-assembles a 7-10 person expert panel tailored to that content and scores it out of 100.

The flow has 7 steps: (1) Intake — what's being scored, type, domain, variants; (2) Panel auto-assembly — content-type experts + 1-3 domain experts + two always-fixed experts: AI Writing Detector (weighted 1.5x, 24 AI-writing patterns) and Brand Voice Match; (3) Rubric selection — content / strategic / conversion / visual / evaluation quality; (4) Recursive scoring loop — target 90+, if the aggregate is low it identifies the top 3 weaknesses, revises, re-scores (max 3 rounds, brutally honest, no padding); (5) Output — winner + score + full iteration trail; (6) Feedback-to-source — if the content came from another skill, a brief that improves the source; (7) Memory — it learns approval/rejection patterns into references/patterns.md and docks points for them going forward.

When to use it

• When asked to "expert panel / score / quality-check this."
• When you need to objectively pick the best of A/B/C variants.
• As a quality gate before shipping a landing page, email, headline, or strategy doc.
• When another skill's output needs a quality gate (e.g. cold-email, copywriting).
• To catch AI-writing smell, brand-voice drift, or known-bad patterns.

Method / frameworks

The engine rests on rubric-based expert-panel calibration: multiple evaluators score the same artifact independently, a weighted average is taken (humanizer 1.5x), and a delta analysis prioritizes the top 3 weaknesses. One of five rubrics is chosen by content type — each with 5-8 core criteria, weights tuned by content type and audience segment (ISO-style multi-criteria quality scoring). The AI Writing Detector uses a 24-pattern detection rubric, flagging text that is systematically "too correct / too regular / too register-stable." Pattern enforcement checks references/patterns.md against the content every round and docks points for known-bad patterns before expert scoring even begins — so known mistakes are penalized even if individual experts miss them. The 90 threshold is a quality gate, not a downstream-performance guarantee; the iteration trail is always shown as part of the value.