Teaching AI Taste

An Amplifier Case Study

185 lines of markdown, zero code

anderlpz/amplifier-bundle-design-intelligence-enhanced

The Missing Half

A bundle can track every design trend and still not know if you'd like any of them

The Design Intelligence Enhanced bundle faithfully gathered what's happening in design. But "what's out there" is only the informed half — taste is the judgment half, and it was missing.

First, the proof that the informed half already worked.

Since 2026-01-13 it pulled design trends every week, on its own

A weekly research recipe fetched RSS feeds into a dated archive, generated a trend summary, and updated a 30-day rolling index. The collection engine was real and running.

Yet all that collecting hit a ceiling.

3
RSS sources: Awwwards, Siteinspire, The FWA
Weekly
Cadence to stay current
30-day
Rolling archive index
44 days
Research shipped before taste
The Ceiling

It only ever got more informed — never smarter

More feeds meant more data, not better judgment. The agents had no memory of your taste and no loop connecting accumulated trends to your personal preferences.

The obvious instinct is to build a feature. It was the wrong instinct.

The Reframe

The fix wasn't a feature to build — it was judgment to write down

Taste turned out to be an instruction problem, not a code problem. Instead of engineering new machinery, the answer was to teach the existing agents how to reason about preference.

Here's the instruction that carried the judgment.

One 185-line markdown file taught the agents to reason in three modes

context/taste-awareness-instructions.md loads and merges your user-global and per-project taste profiles, then reasons in a three-mode protocol. No code — just prose.

And it does more than decide once — it learns.

1
Decide
Apply the known profile directly
2
Flag
Surface a conflict for the user
3
Recommend Evolution
Propose the profile should change

After repeated observations it proposes confidence-graded profile updates

On at least 2 occurrences of a signal, it proposes a profile update, escalating confidence Low → Medium → High. That's the loop from data to remembered taste the bundle never had.

All of it shipped as markdown, in a single day.

≥2
Repeated observation
A preference signal seen at least twice
Profile update proposal
Suggests evolving your taste profile
Confidence escalation
Low → Medium → High

Judgment for 8 agents, in a single day, with zero new code

Registered by one line in bundle.md and six-line frontmatter edits per agent, taste-awareness reached 7 advisory agents plus research-analyst. Across all 5 commits: only markdown and yaml — 0 code files.

The reframe was right — so what's the pattern?

8
Agents given judgment (7 advisory + research-analyst)
0
Code files touched (.py/.js/.ts/…)
1 day
All 5 commits dated 2026-02-26
247
Total lines added across the feature
The Pattern

When a system is informed but not smart, the missing piece is often instructions, not infrastructure

Look for judgment problems disguised as engineering problems. Sometimes the capability you're about to build is one you can simply write down.

Teaching taste took prose, not a platform.

Sources

Research Methodology

Feature status: Shipped

Data as of: repo history 2025-12-23 → 2026-05-20 (49 commits)

Case-study subject: anderlpz/amplifier-bundle-design-intelligence-enhanced (author Alex Lopez), an enhanced fork of microsoft/amplifier-bundle-design-intelligence.

Research performed:

Notes: The 185-line figure is the file size at introduction (commit bbd8291); it has since grown to 204 lines. Primary contributor: Alex Lopez (all 5 taste-awareness commits).

More Amplifier Stories