How I design with AI.
I connect Claude to the tools where product work lives: Jira, Confluence, GitHub, Figma, zeroheight and the Flutter codebase. AI gathers context, drafts and documents. I frame the problem, make the decisions and approve what gets published.
The workflow
Start from the product owner's input.
Claude pulls epics, stories and requirement pages from Jira and Confluence through an MCP server authenticated with my API token. Nothing is copied by hand, so nothing gets lost.
- My role
- Check scope, find gaps and open questions.
- Output
- A complete brief
- Tools
- Jira · Confluence · MCP
I plan the design approach myself.
Problem framing, research methods and which users to involve are my decisions. AI stays out of this step on purpose.
- My role
- Frame the problem, plan the research.
- Output
- Design-thinking plan
- Tools
- Design thinking
Sharpen the prompts, run them with context.
I refine prompts with ChatGPT, then run them in Claude, which already has the Jira and Confluence context loaded.
- My role
- Write, test and iterate prompts.
- Output
- Reusable prompts
- Tools
- ChatGPT · Claude
Work with the real system.
Claude works in local clones of the design-guide and app repositories and uses the GitHub MCP server for issues and pull requests. Component rules come from zeroheight, design context from the Figma MCP server.
- My role
- Decide what the AI may read and change.
- Output
- Real components and rules
- Tools
- GitHub · zeroheight · Figma
Design, then prototype on the codebase.
Designs in Figma, then prototypes built in the Flutter repository and run locally, so they behave like the product. Prototype code stays separate from production; developers own what ships.
- My role
- Design decisions, quality, review.
- Output
- Prototypes that feel real
- Tools
- Figma · Flutter
Documentation ships with the design.
Claude drafts the spec page in Confluence, creates tickets and user stories in Jira and commits prototypes and docs to GitHub. I review each one before it's published or merged.
- My role
- Review and approve.
- Output
- Traceable handoff
- Tools
- Confluence · Jira · GitHub
Connections
What Claude takes over
- Figma designread via Figma MCPJira ticketscreated via Jira MCP
- Design decisionsfrom the working sessionConfluence spec pagedrafted via Confluence MCP
- Requirementsfrom Confluence and JiraUser story draftsreviewed by me
- Prototype and docsfrom the working sessionGitHubcommit and pull request
Triggered by me during a working session, not scheduled in the background. Nothing is published or merged without my review.
Principles
At Bosch I also built an internal AI toolkit for designers: practical prompts for research, ideation and UX writing, plus guidance on where human judgement stays essential. Bosch case study →
From thesis to practice
In 2019 my bachelor thesis described three stages of access an assistant needs to be truly useful. MCP now provides that access in a standard way.
External access
Open, send and manage files and apps.
Today: Claude opens repos, pages and tickets.
Internal access
Understand the content inside files and apps.
Today: Claude reads requirements, specs and the design guide.
Full access
Process content and actively support the user.
Today: Claude drafts specs, tickets and prototypes. I decide.
How this site was built
- Read the old site: 10 pages, jQuery, Bootstrap 3, seven stylesheetsAI
- Interview on positioning, workflow and impactMe
- Research public figures: Bosch, Porsche, DS CoreAI
- Compress 163 MB of images and video to about 15 MBAI
- Art direction: palette, typography, references, more white spaceMe
- Design system, templates, English and German, small Python generatorAI
- Review and approve every claimMe