Design Intelligence for AI-built interfaces
Developers can hand an AI that one prompt and get something usable. But expertise, not the ask, decides whether the result looks like every other dashboard on the internet — or like your vision.
The whole question this deck resolves: generic, or yours?
This isn't a vague promise. The Design Intelligence bundle ships seven design advisory agents — each a real agent file with a distinct, verifiable size.
Real expertise exists. So why does the AI still go generic?
art-directordesign-system-architectcomponent-designerlayout-architectanimation-choreographerresponsive-strategistvoice-strategistAsk for a healthcare dashboard and it researches other healthcare dashboards — then returns a "competent remix of existing work." Low innovation, because it inherits the category's assumptions.
To break out, research itself has to change.
Instead of copying the category, the bundle prescribes a 20% category / 50% purpose-driven / 30% divergent blend — extracting Level 3–4 interaction and cognitive patterns, not just visual ones.
Fresh-but-grounded. But what keeps that judgment consistent?
Taste is stated and shared, not improvised. Agents reason through the Nine Dimensions and Five Pillars, so advisory judgment stays consistent and defensible.
Grounded judgment sets the bar. Now — who enforces it?
Quality stops being a hope. design-check validates the work against the design system — hardcoded colors, wrong spacing, missing states, AI-fingerprint patterns — as silent agent self-correction.
Which is why the payoff feels like magic.
The 9.5/10 standard is enforced silently before the work ever reaches you. What's left is the part only you can do — the final 5%.
The bar is cleared quietly, and the result is still yours.
"AI handles execution; you provide sensibility. The system ensures quality; you provide meaning."
The machine enforces the standard as a process; you supply the sensibility that makes it not every other dashboard.
Primary source: colombod/amplifier-bundle-design-intelligence-enhanced (fresh gh clone: origin = colombod, upstream = anderlpz fork; README install line references microsoft/amplifier-bundle-design-intelligence).
Feature status: Phases 2.0–2.3 marked Complete (bundle.md version 2.3.0); phase 3.0 Feedback Planned, not shipped.
Timeline: initial commit 2025-12-23; latest 2026-05-20; 49 commits (git rev-list --count HEAD). Single git tag v2.1.0 despite bundle.md declaring v2.3.0.
Research performed:
for a in art-director design-system-architect …; do wc -c agents/$a.md; donels agents/*.md | wc -lgrep -rn '9.5' . --include='*.md' --include='*.yaml' | grep -v '.git/' | wc -lgrep -rl 'Nine Dimensions' agents/ context/; grep -n 'Five Pillars' README.md context/philosophy/*.mdsed -n '90,108p' bundle.md; grep -n 'Layer' agents/design-check.mdgit shortlog -sn --all — Alex Lopez 44 (~90%), Brian Krabach 5 (49 total).Gaps: No quantitative benchmark proving output empirically reaches 9.5/10 was found — 9.5/10 is an explicit design standard/target repeated across docs, not a measured result. The 20/50/30 blend and Level 3–4 patterns are documented prescriptions, not measured outcomes. Canonical org attribution (colombod vs anderlpz vs microsoft) could not be independently disambiguated beyond git remote output.