Medium: How AI became part of a product designer's UX workflow at Cobalt

Laís Lara Vacco is a product designer at Cobalt, a B2B cybersecurity platform. Her July 2026 article in Medium’s Bootcamp publication is a stage-by-stage account of how AI entered her daily work — not as a single tool swap, but as a slow, experimental layering across six distinct workflow phases.

What the article covers

Stage 1 — Learning and research. Vacco uses NotebookLM to summarize internal documentation and generate audio briefings on complex features. For open questions, she routes to Claude or Gemini rather than interrupting engineers. She also processes meeting transcripts through AI follow-up questions to extract action items.

Stage 2 — Understanding product logic. Cobalt’s permission model is complex. Vacco uses Claude Code to read and reason about GitHub repositories, letting her understand engineering decisions independently before validating assumptions with developers. This reduced back-and-forth while keeping her interpretations grounded.

Stage 3 — Problem-solving and concepting. She uses AI to generate ASCII diagrams and workflow maps for scenarios too complex to sketch quickly. She also applies the BMAD Method, which assigns specialized agent roles (UX Designer, Product Manager) to surface contradictions and edge cases before screens are designed.

Stage 4 — Prototyping. Vacco moved from static wireframes to clickable AI-generated prototypes early in the process. She tested Claude Code, Subframe, and AI Studio, ultimately settling on Subframe as her primary tool for AI-native prototyping aligned with design systems.

Stage 5 — Design system integration. This is the article’s most original contribution: Vacco created a DESIGN.md file — a structured Markdown document that teaches AI tools about Cobalt’s design patterns, spacing rules, component guidelines, and naming conventions. She iterated on this file using evaluation prompts to test how well AI could apply the documented rules. When the output improved, the documentation had done its job.

Stage 6 — Design-to-code bridge. She uses Subframe as a bridge between design and engineering, generating component code that conforms to the design system rather than generic HTML output. Tools like Builder.io and Figma Make were tested but rejected for reasons specific to her workflow.

Who benefits from this article

The article is most useful for product designers at companies with an established design system who want to reduce AI output misalignment. The DESIGN.md concept is directly applicable: any designer who has spent time correcting AI-generated components that ignore spacing or color tokens will recognize the problem it solves. Vacco is candid about what did not work — AI Studio, Builder.io — which makes the article useful for teams choosing between tools rather than just following a success story.

The article also touches on team culture: sustainable AI integration requires treating experimentation as part of regular work, not a side project.