Rebuilding Our Workflow with Generative AI

Relevant Case Study: Bridging the Gap: Design to Development at Treatspace, Inc.

When I moved into product design at Treatspace, project requirements were not a thing. We would discuss ideas as a group, talk through a game plan, and break out. Scope creep was constant and unexpected technical requirements often forced reworks for both design and engineering.

Using a combination of Slack, Jira Product Discovery, and Claude, I was able to create an automated workflow that provided the product team with clear project requirement documents for each new idea that spawned from a Slack conversation – Slack being the space where most of our product discussion took place.

These PRD’s were not only clear and concise references for all teams to agree on, but they were formatted to feed directly into our generative AI tools. This provided a nearly automatic Idea-to-Prototype workflow that allowed us to validate real, working features, rather than a half-baked idea.

Leads within the Product Team adjust the generated PRD’s manually to provide additional context and validate timelines. Once the document is approved, the feature/idea is generated using our design system and codebase. This results in a rough deliverable that can be adjusted by the design team, reviewed by the engineering team, and tested by the customer success team.

This process would take anywhere from days to weeks before generative AI – things needed to be perfect due to the time invested in making working prototypes, and missing the mark on a user test was devastating. Now user tests are something we can always look forward to, and the lessons we learn from them don’t sting like they used to.

There will never be a replacement for 1-on-1 human user testing sessions, but they can sometimes be more effort than they are worth. 1-on-1 sessions are still an option when necessary, but we also started doing recorded solo sessions. These sessions are recorded and transcribed, allowing for many tests at once. Each session is summarized and ultimately aggregated into a master list of common pain points and bugs.

Sessions can be initiated manually by our team, or automatically by users themselves on the platform.

This is pretty much where the humans take over. We get a final prototype, adjusted for feedback, and begin reviewing the output. We smooth out the rough edges and make sure the engineers have a useful figma file. Often times font styles will not be applied, or a color may not have the correct variable – we correct these mistakes and adjust the design to meet PRD success criteria/design best practices.

Although this is a human endeavor, AI aids us along the way with Claude & Figma Agents being able to audit the initial prototypes and make bulk adjustments.