From ambiguity to production.
I join before a task is ready for engineering: map the scenarios, build a prototype in Figma or code, write down the states and tokens, and see it through to production.
I currently run this loop on the design system at Yandex HR Tech (previously across design systems at Sber).
How the work goes
(4)Context
The problem, scenarios, constraints, and how we will know it works. Without that, a mockup only hides the ambiguity.
Prototype
Not the whole product — just the scenarios where the design might break. Built in Figma or directly in code, whichever exposes the edge cases faster.
Rules
States, components, tokens. One specification for design, engineering, and coding agents so nobody invents a parallel version.
In production
A mockup is not the finish line. I stay through implementation, browser checks, and linting rules so the system holds together over time.
With an agent
The agent handles the draft work step by step: checks what already exists in the system, assembles a first pass, and runs the checks. I scope the task, set the constraints, and review each step.
One giant prompt for the whole job doesn't work: the model writes the code and grades its own homework. Generation and verification have to be separate steps.
The product decision, the scope tradeoffs, and anything that ships under my name stay with me.
A good process does not promise the model will never fail. It catches the failure before it quietly ships.
More on this in the essay on how AI changes processes.