Design Thinking

AI as Design
Thinking Partner

A single brainstorming session that redefined how we think about building with AI

February 2026
Active -- Shipped as amplifier-chat
The Shift

What if AI could help you think,
not just build?

We ask AI to write code, fix bugs, and execute plans. But what about the messy, creative work that happens before the first line of code?

The Problem

The ideation gap

Design in Isolation

Product vision starts as a solo exercise -- sketching on whiteboards, drafting docs alone, then scheduling reviews days later to get feedback.

AI as Executor Only

Current AI workflows excel at implementation -- "build this component" -- but rarely participate in the upstream thinking: "what should we build and why?"

Shallow Conversations

Most AI interactions are transactional. A prompt, a response, done. Deep collaborative exploration -- the kind that produces breakthroughs -- is rare.

The Experiment

One developer. One question. One sustained conversation.

A developer sat down to design the next generation of Amplifier's user interface -- a web-based replacement for the CLI. Instead of sketching alone or scheduling a design review, they opened an Amplifier session and started a conversation.

Developer
I'd love to see dynamic UI and leverage the model to figure out the UI for me. Not just build it -- design it.
Amplifier
That's a fascinating direction. Let's explore what "generative UI" means in this context. If the model decides not just the content but the layout and visual composition...

What followed was not a prompt-response exchange. It was a sustained exploration.

The Arc

The arc of a brainstorming session

Opening
Generative UI Exploration
What if the AI decides not just content but layout, rendering, and visual composition? Exploring dynamic interfaces that reshape themselves based on context.
Middle exploration
Conversation-Centered Design
Layouts where conversation is the primary interface -- faded history with scroll-back, collapsible tool outputs, progressive disclosure of complexity.
Research enrichment
Multi-Agent Research Enrichment
While the creative conversation flowed, Amplifier dispatched research agents to pull in competitive context, UX patterns, and user mental models -- in real time.
Convergence
Philosophy Crystallization
Converging from exploration to conviction. The session produced a product principle: "We don't worry about the token. We worry about the experience."
Concept 1

Generative UI: AI designs the interface

The first breakthrough idea: instead of building a fixed UI that displays AI responses, let the AI dynamically decide how to render each response.

Traditional approach

Developer designs the UI. AI fills in the content. Every response looks the same -- a text block in a chat bubble, regardless of what's being communicated.

Generative approach

AI decides the layout per response. A code review gets a diff view. A comparison gets a table. An architecture discussion gets a diagram. The container adapts to the content.

The UI becomes a blank slate, not a template
Concept 2

Conversation as the primary interface

The session explored a radical simplification: what if the conversation IS the application? Not a sidebar to some other tool, but the center of the experience.

Progressive Disclosure

Tool outputs collapse by default. Thinking blocks show a preview. Complexity reveals itself only when the user asks for it.

Faded History

Older conversation turns fade visually, creating a natural "focus on the present" while keeping scroll-back available. Context without clutter.

Streaming as Presence

Real-time token streaming isn't just a technical feature -- it creates a sense of the AI thinking alongside you. The stream IS the collaboration.

The Differentiator

Research agents enriching the brainstorm

While the creative conversation flowed, Amplifier dispatched multiple research agents simultaneously -- pulling in context that no single person could gather that fast.

Competitive Analysis

Surveyed existing AI chat interfaces, identifying patterns and gaps in the market

UX Pattern Research

Retrieved conversation UI patterns, progressive disclosure approaches, and information density studies

User Mental Models

Analyzed how developers think about chat vs. IDE vs. terminal interfaces

Architecture Patterns

Investigated protocol designs for real-time streaming, WebSocket vs. SSE tradeoffs

Research happened during the brainstorm -- not before, not after. Context arrived as the thinking evolved.

The Breakthrough
"We don't worry about the token that is used. We worry about the experience."
Emergent product principle -- late in the session

This wasn't pre-planned. It crystallized from sustained, uninterrupted exploration -- the kind of insight that only emerges from deep, uninterrupted collaborative thinking.

The Distinction

Thinking partner vs. execution partner

AI as Executor
Takes instructions, produces output
Optimizes for correctness
Session ends when task is done
Answers questions
Works within given constraints
AI as Thinking Partner
Explores problem space together
Optimizes for insight
Session deepens through sustained exploration
Asks better questions back
Challenges and expands constraints
The Outcome

From brainstorm to shipped product

The brainstorming session didn't just produce ideas. It produced a vision clear enough to become a 5,184-line specification -- and then a shipped product.

Brainstorm
session
5,184-line
NL spec
amplifier-web
prototype
amplifier-chat
shipped

The Spec

5 documents covering product context, protocol, backend, frontend, and implementation details. Created Feb 13-17, 2026 in the amplifier-web-spec repo.

The Prototype

First implementation as amplifier-web -- React/TypeScript frontend with WebSocket protocol. Validated the core concepts.

The Ship

Evolved into amplifier-chat -- a Preact SPA using SSE streaming. 163 commits, actively developed.

Development Velocity

What one brainstorming session produced

1
Brainstorming session
5,184
Lines of specification
163
Commits shipped
8,942
Lines of UI code

Primary contributor: Samuel Lee (145 of 163 commits in amplifier-chat) · Paul Payne (9 commits)

Spec: 6 commits across 5 days (Feb 13-17, 2026) · Implementation: March 3-20, 2026

The Bigger Picture

Why this matters beyond this project

AI moves upstream

The most valuable moment in any project is the beginning -- when assumptions are formed and direction is set. If AI can participate here, it changes everything downstream. Not just faster execution, but better decisions about what to execute.

Depth over breadth

This wasn't about efficiency. It was deliberation. The value wasn't in getting a quick answer -- it was in the late insight that the opening question couldn't have reached. AI as thinking partner requires patience and trust in the process.

Research-enriched creativity

The multi-agent research pattern -- dispatching agents for context while the creative conversation continues -- is a new workflow. It's like having a research assistant who fetches references while you sketch on the whiteboard.

Experience as north star

"Experience over tokens" emerged from the conversation, not from a planning document. The best product principles are discovered through exploration, not decreed from above. AI can be part of that discovery process.

Sources

Research Methodology

Data as of: March 22, 2026

Feature status: Active (amplifier-chat is actively developed and deployed)

Research performed:

Primary source for brainstorming session details: Developer's firsthand account of the 27-turn session. Exact conversation logs were not available for independent verification. Specific quotes and turn-by-turn arc are from the developer's recollection.

Repositories: All repos under microsoft/ org -- amplifier-web-spec, amplifier-web, amplifier-chat

Gaps: Exact conversation transcript not independently verified. Turn numbers for specific insights are approximate from the developer's account. Multi-agent research dispatch patterns are described conceptually -- specific agents and queries from that session were not logged.

Try It Yourself

Start with a question,
not a task.

The next time you're designing something new, open an Amplifier session and explore. Don't ask it to build. Ask it to think with you.

amplifier-chat

Try the web UI that emerged from this session

amplifier-web-spec

Read the 5,184-line spec that captured the vision

Your next project

Use AI for the thinking phase, not just the building phase

More Amplifier Stories