COMPANY
Kustomer
ROLE
Head of Product Design
YEAR
2024
Impact Numbers
CONCEPT → WORKING PROTOTYPE
Time to testable direction
STATIC SPECS REPLACED
CHALLENGE
AI made engineering faster than design could feed it. Handoff became the bottleneck, and quality decisions fell out of the loop.
APPROACH
Replace specs with prototypes, reviews with live AI interactions, and one-off critique with reusable checks and guardrails, built with engineering rather than handed to it.
OUTCOME
I built an AI-native operating model that cut time to a testable direction by 80%, moved prototyping from three days to three hours, and replaced 75% of static specifications.
She codified the whole thing into reusable markdown-based skills so AI stopped being individual experimentation and became a shared operating system. Then she moved critique from vibes to observable output quality.
Arkady Sokolov
Product, AI | ClickUp
AI reviews working prototypes against shared UX standards before release.
# AUDIT UX: {FEATURE}
## What we encode
Undefined states
Dead ends
Abandonment risks
Recovery failures
Broken user journeys
## OUTPUT
Prioritized findings tied to specific routes, screens, and states.
Version-controlled rules give humans and AI the same system to build from.
## What we encode
Accessibility requirements
Content standards
Component usage
Brand voice
Error handling
Interaction patterns
## Output
Reusable AI instructions and testable implementation constraints.
A decision rule for where automation ends and human judgment begins.
## Decision tree
- Reversible / Low-Stakes → model decides, log it
- Reversible / High-Stakes → model proposed, human confirms
- Irreversible → human decides, model assists
## Default
Expose rationale, sources, confidence, and an undo path.
Review the AI behavior, not just flows and screens
## What we critique
- Outputs → are responses useful and correct?
- Recovery → what happens when the model fails?
- Confidence → can users trust the result?
- Boundaries → where does automation stop?
## Output
Evaluation criteria, prompt and UI changes, guardrail revisions, and new design rules.
WORK IN LIVE SYSTEMS OVER STATIC SPECS.
01
From static handoffs to LLM prototyping, real outputs, and same-day iteration.
02
Shared skills for prompts, component rules, output evals, and HITL guardrails.
03
Critique against real use cases and edge cases before shipping.
04
Design and engineering define behavior and review outputs together.
Work in live systems over static specs.
Evaluate AI against real use cases.
Turn experiments into a shared OS.
Humans remain accountable.
hello@ydoll.com
At Kustomer, I rebuilt the design operating system around a new reality: designers, engineers, and AI systems working in the same loop.
AI made engineering faster than design could feed it. Handoff became the bottleneck, and quality decisions fell out of the loop.
Impact Numbers
CONCEPT → WORKING PROTOTYPE
Time to testable direction
STATIC SPECS REPLACED
Impact Numbers
CONCEPT → WORKING PROTOTYPE
Time to testable direction
STATIC SPECS REPLACED
She codified the whole thing into reusable markdown-based skills so AI stopped being individual experimentation and became a shared operating system. Then she moved critique from vibes to observable output quality. Then she cut time to decision by 80% because why not
Arkady Sokolov
Product, AI | ClickUp
I identified a recurring gap in our AI workflow, built the first reusable UX checks myself, and scaled them into shared standards for
Design and Engineering.
AI reviews working prototypes against shared UX standards before release.
# AUDIT UX: {FEATURE}
## What we encode
Undefined states
Dead ends
Abandonment risks
Recovery failures
Broken user journeys
## OUTPUT
Prioritized findings tied to specific routes, screens, and states.
Version-controlled rules give humans and AI the same system to build from.
## What we encode
Accessibility requirements
Content standards
Component usage
Brand voice
Error handling
Interaction patterns
## Output
Reusable AI instructions and testable implementation constraints.
A decision rule for where automation ends and human judgment begins.
## Decision tree
- Reversible / Low-Stakes → model decides, log it
- Reversible / High-Stakes → model proposed, human confirms
- Irreversible → human decides, model assists
## Default
Expose rationale, sources, confidence, and an undo path.
Review the AI behavior, not just flows and screens
## What we critique
- Outputs → are responses useful and correct?
- Recovery → what happens when the model fails?
- Confidence → can users trust the result?
- Boundaries → where does automation stop?
## Output
Evaluation criteria, prompt and UI changes, guardrail revisions, and new design rules.
Work in live systems over static specs
Design prompts, outputs, rules, and recovery paths, not just flows.
Evaluate AI outputs against real use cases.
Co-create with engineering from the start.
Use real data and real user signals as early as possible.
I rewired my own practice first, then used that learning to help the team move from static design artifacts to AI-assisted prototypes, research synthesis, product critique, and shared evaluation systems.
Cursor
Claude Code/Design
Code-based prototyping
AI-assisted research synthesis and product analysis
Moved design team from static handoffs to AI-native co-creation: prototype with LLMs, test real outputs, iterate in hours.
Introduced Markdown-based skills to standardize AI usage across prompts, component rules, output evals, and human-in-the-loop guardrails.
Turned AI from individual experimentation into a shared operating system.
Moved design critique from subjective review to observable output quality. AI-generated work is evaluated against real use cases, defined edge cases, and documented product intent before shipping.
Eliminated traditional handoff by embedding design and engineering in the same loop: co-defining behavior, prototyping together, and reviewing outputs collaboratively.
Compressed idea-to-working-output timelines from days or weeks to hours, reduced interpretation gaps and rework, and tested design quality against real conditions earlier.
Creative thinking did not disappear. It moved from static files into system behavior: prompts, rules, evaluation criteria, and the user’s voice in the process.
