4.0 KiB

Raw Blame History

Lesson 4: The Feedback Loop

Module 19: Design Space | Time: 10 min

How Agents Learn Taste

When you work with a designer and they suggest improvements, that's not just a correction — it's a preference signal. The feedback loop captures these signals as linked pairs, and over time, the agent develops design taste.

Philosophy: The feedback loop captures solutions, not complaints. The "before" state is context. The "after" state — the improvement — is the real knowledge.

The Flow

Agent creates a design
    ↓
Designer suggests an improvement
    ↓
Agent captures BEFORE (the starting state)
    ↓
Agent asks: "What would make this better?"
    ↓
Designer explains (or agent infers)
    ↓
Agent applies the improvement
    ↓
Agent captures AFTER (the improved version)
    ↓
Both saved as a linked pair
    ↓
Agent confirms: "Learned: [X] works better because [Y]"

The WHY Question

This is the most valuable moment. The designer's reasoning is what makes the learning transferable.

Ask naturally — don't interrogate:

Forward-looking: "What would make this feel right?"
Specific: "Should it be more open / minimal / bold?"
Outcome-oriented: "What feeling should this create?"
Inference: "Got it — lighter weight works better here because [reason]. Right?"

Sometimes the designer can't articulate why. That's fine. Capture the observable change: "Improved from bold to light weight — designer's intuitive direction. The result creates a calmer, more elegant feel."

Framing Matters

How you frame the learning determines whether the Design Space becomes a library of solutions or a list of complaints.

Good Framing (solutions)

"Light heading weight (300) creates elegance — works better than bold for confident calm brands"
"80px section padding gives content room to breathe — outperforms 48px on service pages"
"Left-aligned text follows natural reading flow better than centered for body copy"

Bad Framing (complaints)

"Designer hates bold headings"
"48px padding was wrong"
"Centered text is bad"

The good framing is actionable. The bad framing is a dead end.

Capture Format

capture_feedback_pair({
  before_description: "Hero section with H1 at 48px bold (700) Rubik,
    navy background, full-width. Bold heading feels authoritative
    but heavy.",
  after_description: "Hero section with H1 at 48px light (300) Rubik,
    navy background, max-width 800px. Light weight creates elegance
    and breathing room. Same authority, less weight.",
  reasoning: "Bold headings feel corporate and generic. Light weight
    at large sizes is distinctive — the brand is confident calm,
    not loud authority.",
  pattern_type_before: "rejected",
  pattern_type_after: "approved",
  project: "whiteport",
  topics: ["typography", "heading-weight", "brand-voice", "elegance"],
  components: ["hero-banner", "heading-h1"]
})

Both descriptions should be specific enough that someone could recreate the design from the text alone.

The Learning Curve

Stage	Pairs	Agent Behavior
Cold start	0-10	Individual solutions. "Light headings work better for this brand."
Accumulation	10-50	Principles emerge. "Understated elegance across typography, spacing, color."
Taste profile	50+	Agent anticipates improvements. "The lighter option with more whitespace will work."
Design DNA	100+	New agents inherit design sensibility from day one.

The cold start is unavoidable. But every feedback pair accelerates the learning. By project 3-4, agents start making noticeably better first proposals.

Key Takeaway

The feedback loop isn't an interruption to design work — it is the design work. Every improvement you suggest teaches the system what good design looks like. Over time, the system learns to produce it.

← Lesson 3 | Next: Lesson 5 →

4.0 KiB Raw Blame History