2026
muddyhands
A voice-first reflection tool for potters — and a study in what an AI-native design workflow can look like.
Role
Product designer
Tools
v0
Figma+Figma Make
Figma MCP + Cursor + Claude
Duration
5 days + iteration
Official Website
muddyhands
Built a voice-first reflection tool for potters in 5 days. Then rebuilt it through an AI-native workflow with Figma MCP, Cursor, and Claude.
The workflow is the part that matters.
Capture now. Reflect later.
Design the workflow that makes both possible.
ORIGIN
Pottery class. Hands covered in clay. The thoughts worth keeping happen exactly when you can't write them down. Stop to capture and you break the rhythm. Don't stop, and the thought is gone. Anyone doing skilled physical work hits this — surgeons, chefs, musicians. The moment of insight is what the interruption would cost.
Pain points are:
01/
Insight happens during making.
Stopping to capture it breaks the work.
The acute problem
02/
Without a record, reflection is shallow.
Progress stays invisible over time.
The chronic problem
How migh we
help people in physical creative practices document their process — without interrupting the work itself?
The Product
Built in 5 days. Tested in a pottery studio. Live.
TWO KEY FEATURES
From user journey map, I discovered that capture and reflection are different cognitive modes, this lead to MuddyHands' two features for the process, not one. Meanwhile, given the time limit of 5 days, I elimated other possible features to secure the scope before next step.
01/
capture
Capture mode is one-tap voice. Minimal text. Designed to disappear into making.
02/
reflection
Reflection mode is structured. Sessions, pieces, notes. Designed to invite slow review when attention is free.
NOT IN SCOPE
User profile setting
NOT IN SCOPE
Cross-session analytics
NOT IN SCOPE
Image capture
NOT IN SCOPE
Social sharing
INFORMATION ARCHITECTURE
With the scope locked, I designed the information architecture around cognitive state.
Designing the IA this way surfaced a key insight: Reflection works bottom-up, not top-down.
Users need to understand individual pieces before making sense of a session. The IA enforces this order by design, preventing premature summarization or false completion.
TWO INTERACTION PROBLEMS SOLVED
When sketching the possible flows, I noticed some confusing spots.
Actions vs States
"Complete" was doing two jobs — user action and system status. I split them. Done ends capture. Active and Completed describe state.
Optional reflection
I removed mandatory review. Reflection became self-directed. That decision changed the emotional tone of the whole product.
DESIGN SYSTEM
Before any workflow question could be answered, the design system had to exist. Not as a style guide — as structured ground truth.
Warm palette. Typography that breathes. Components for status, sessions, pieces, capture, reflection. Not aiming for completeness — aiming for hierarchy, calm tone, and unambiguous status signaling.
USER TESTING
I tested the MVP with two potters from my weekly class.
from
Josie, my classmate
"I usually forget what I did with each piece after class. Having it organized like this would make it much easier to look back and remember."
Josie felt the app would make it easier to keep track of pieces and processes across a session. She mentioned that having notes organized by piece would help her reflect later on what worked and what didn't.
The sprint shipped. But the system wasn't really finished. It was a brief. The real test was rebuilding it through a workflow that scaled.
The AI Workflow Evolution
Three stages, three different relationships with AI
Flow validation
v0 answered the architecture question in hours.
But v0 reinterpreted my spacing and hierarchy. It had opinions. I needed execution, not interpretation.
V1 implementation
Figma as source of truth, Figma Make for execution. MVP live by Day 3. End of sprint.
Intergrated rebuild
Claude with direct read on my Figma file. Real tokens, real components, real ground truth. Designing alongside AI, not negotiating with it.
what happend here
CONTENT DESIGN AS CODE
I built a Claude Code command called tone-system — markdown defining voice, status tone, deletion feel, words out of scope.
Slash command - Claude executes against it.
Voice enforced by the system, not by my willpower.
what happend here
TONE IN PRACTICE
The empty state, when no session exists, says:
"Nothing yet. Begin when you're ready."
Most apps push Get started. Pressure — exactly what this product removes. For potters, the empty state isn't a problem to fix. It's where you are before you begin.
what happend here
WHERE IT BROKE
Claude misread my capture trigger as a nav bar — because it looked like one.
Fidelity drifted when context got too long.
Same lesson both times: AI defaults to the most common pattern. My job is to define the exception clearly enough that AI can hold it.
On a solo project, a tone-system command is one file. On a team, it becomes a shared content design layer every designer and agent references.
The command library is the beginning of something larger.
Final Takeaway
lesson learned about the product
strong ideas don't need long timelines,
once the problem is clear, ship the core behavior fast enough to learn from real users.
reflection on the workflow
AI is most useful when it works from the same ground truth I do — tokens, components, tone rules. Not natural language. The job of design in an AI-native workflow is to
author the ground truth clearly enough that AI can execute against it.
looking forward
how do you design this for a team
— moving fast, across multiple products, with designers at different levels of AI fluency?