This project never had a specification. It had sentences like "I want to recreate these two screenshots", "the zoom doesn't quite work", and "could we get a total tile for small multiples?" — each was answered not with a document, but with a rendered image to look at, judge, and refine. Describe, render, look, refine: that loop replaced the spec, the backlog, and half the development process. This session shows how far it carries you — all the way to production Power BI custom visuals.

The tooling is deliberately boring: Claude Code as it ships — no fine-tuning, no custom AI infrastructure — plus the normal dev toolchain (Node.js, the pbiviz tools, a headless browser). The only "AI setup" that accumulated along the way was a set of plain-markdown skill files the assistant wrote for itself as it learned the project.

The path has three stages, and none of them starts in the visuals SDK. First: plain HTML prototypes in the browser, where chart logic is ten times faster to iterate on — the prototype becomes the spec. Second: Deneb and Vega-Lite, where declarative templates carry you surprisingly far, until interactivity shows you the ceiling. Third: a real TypeScript .pbiviz — where a headless-browser render harness keeps the look-and-refine loop down to minutes, and real-mouse-click tests turn "should work" into "works".

And this is not slideware: we'll run one example live on stage — from the first sentence to a working prototype. Feel free to open your laptop and code along; everything used in the session is MIT-licensed and public, so you can replay the whole path at home.

The examples prove the approach is standard-agnostic: IBCS-style business charts on one end — the standard serving as a borrowed, ready-made design spec — and an interactive project Gantt chart (collapsible phases, dependencies, milestones) on the other. Your target can be a formal standard, a community template, or your own idea.

We close the loop with an AI-assisted security review as a final gate before distribution through the organizational store — being honest about its limits: it raises the bar, but it's no guarantee and no substitute for your organization's review process.

Prerequisites: Solid Power BI report-building experience — no AI or TypeScript experience required. Laptop optional: bring Claude Code and Node.js pre-installed if you want to build along; a public repo provides the rest.

Target audience: BI developers, Power BI leads weighing build-vs-buy, architects and governance leads rolling out visuals safely — and anyone who can judge a chart by looking at it.