A customer feature request used to eat a quarter. The CSM pings the product channel Tuesday at 4pm asking for something specific, and the old loop is triage, rank it against 40 other items, write a PRD, get sign-off, file tickets, build, ship. Eight weeks if you are fast. By the time the customer sees it, their team has hacked around the problem with a spreadsheet. The Instant Prototype Agent replaces that first stretch with a clickable artifact the same day.
Wire it to the right sources
The agent is only as good as its context. Connect the signal channels where requests land: Slack, Zendesk, Gong, and Salesforce. Connect GitHub for your source code and Figma for your design system. Wire Claude Code for generation, Vercel for the preview deploy, and Linear and Notion for the tracking artifacts. Model Context Protocol connects everything.
Five moves, in order
A webhook fires when a request lands in any tracked channel. The agent ingests the full context: who asked, account tier, history of similar requests. It runs a semantic search over the last 90 days of tickets, calls, and threads, because the same request usually has come up before (four times in 90 days is typical). If it is a first-ever request, it flags it for PM review before generating.
Then it generates. Claude Code drafts the feature with your design system, source repo, and production analytics events in context: a new component, wiring, basic state, a realistic data fixture. Three to five minutes. The code commits to a feature branch, a Vercel preview deploys, a draft GitHub PR opens, a Linear ticket files with the customer context, and a Notion doc drafts with the request history. Last, it posts to Slack with the preview URL front and center so the CSM can forward it in one click.
What stays human
The PM decides which requests are worth prototyping at all. The agent flags first-time requests for approval before generating. The PM reviews the PR and decides to merge, iterate, or scrap. The customer decides whether the prototype is close to their real problem. Engineering decides how to turn it into production code. Everyone doing PM, CS, and engineering work still does it, just on a 7-minute cycle instead of an eight-week one.
The compounding effect is the surprise. Once customers know you respond with prototypes, they describe their problems differently: more specific, more concrete, more willing to engage.
Pick one surface where requests land regularly. Wire a webhook to a Claude Code project with your design system and source in context. When the next request lands, trigger the generation manually and time it. Send the preview URL back the same day and notice how different the conversation is.