AI workflow
I use AI every day. I don’t hand it the problem and wait for a finished answer.
It gets me from an idea to something testable much faster. What the evidence means and what ships stays my call.
See the pipelineThe pipeline
- Idea, a question worth testing.
- Lovable or v0, a fast disposable direction.
- Figma, design decisions and the token system.
- Claude Code builds from Figma MCP, Codex runs parallel branches.
- Review, every merge gets read.
- Ship, on Vercel.
Stage by stage
-
First direction
Lovable, v0
When an idea is still fuzzy I throw it at Lovable or v0 and get a clickable concept in hours. An onboarding flow or a 0 to 1 concept becomes something people can react to before anyone falls in love with a direction. These prototypes are disposable and most of them get deleted without regret.
-
Design decisions
Figma
The design system lives in Figma as tokens, styles, and variables. Flows, hierarchy, and the actual product thinking happen here. Nothing about this stage is automated, and that is the point.
-
Research synthesis
Claude
At dhosting I had hours of interview recordings and stacks of open survey answers. Claude did the clustering in an afternoon instead of a day. Reading the clusters, deciding which ones mattered, and turning them into design criteria stayed my job.
-
Build
Claude Code, Figma MCP
Claude Code reads the design system straight from Figma through MCP and builds against a CLAUDE.md file with my spacing scale, typography, and implementation rules. Change a token in Figma and the change shows up in code. This site went from Figma to production Astro exactly that way.
-
Parallel work
Codex
When one session is not enough I run Codex as a second agent on parallel branches. One builds a page while the other cleans up components. I review both before anything merges.
A real build, this site
The visual direction started in Figma. Claude Code built the Astro implementation against my tokens and rules, and I read every change before it merged. Plenty of generated output never survived that read. What you are looking at is the part that did.
What stays mine
- Framing the problem
- Choosing the decision criteria
- Responsibility for what ships
One working rule on top. Confidential client material stays out of unapproved tools.
“Jakub is one of the few people I’d want on a team again without hesitation. He moves between strategy and craft without dropping either, and he uses AI to support good product thinking rather than replace human judgment.”
Where this leaves the work
AI gets me more options, sooner, in testable form. Judgment is still the job.