From a Slack request to a finished Figma design in minutes. An AI agent I researched, sold to leadership, and built — AI fluency turned into a system that scales design across the org.
the pipeline — one front door, a queue, and an agent that ships to figma
Pan (internal codename Kreacher) is an autonomous design agent. A designer or PM files a request in Slack; Pan opens a Jira ticket, picks it up, spawns a Claude subprocess to do the work, and pushes a finished design to Figma — in 5 to 10 minutes. Then it keeps iterating from Jira comments, the same loop a human designer runs, at machine speed.
I put forth an internal research study to understand how designers, product managers, and front-end engineers were actually using AI — and where the biggest pain points were. I synthesized the findings into an opportunity/solution tree to focus on the area with the most impact: system misalignment. Everyone was building their own bespoke AI tools, and none of it connected. I sold that framing to leadership — influencing without owning the roadmap — and got the mandate to build.
Pan lives where the work already happens. A designer opens a Design Request workflow in Slack, names the task, picks a fidelity, drops in a Figma link and the PRD, and describes what they need. No new tool to learn — just a form in a channel.
Every request becomes a Jira ticket. That's deliberate: Jira is the work-tracking and queuing system, which gives the whole thing built-in flexibility. It's also where future "human-in-the-loop" rules will live — flagging requests that are missing tasks or carry risk for a designer to review before Pan attempts them.
Pan polls Jira for open tickets every three minutes during work hours. When it finds one, it spawns a Claude subprocess, does the design work directly in Figma, and reports success back with a link — then goes right back to polling.
Minutes later, a real, on-system design is waiting in Figma. Here Pan built an admin dashboard for AI token consumption straight from the PRD — metric cards, usage breakdowns, allocation controls, alert thresholds, and a cost table.
The first design is a starting point, not the end. Leave a comment on the Jira ticket and Pan picks the feedback up and revises. Here I asked for a token-budget progress bar with a reset date — and it delivered.
Instead of jumping straight to a coded prototype, all generated work lands in Figma. That's intentional: teams that depend on Figma as their source of truth — accessibility engineering, localization, marketing, support — stay in the loop instead of being cut out by an AI shortcut.
Still in its early phases, Pan has shown great promise. Designers and product managers use it to produce wireframes for PRDs and make quick revisions to higher-fidelity designs — reclaiming hours of production work for higher-leverage thinking.