Build a Tableau Viz Extension with AI, Live and No Code

Eric has built dashboards entirely in native Tableau and entirely as vibe coded web pages. Viz extensions are the middle, leaning Tableau: the chart is a web element, but it runs inside Tableau's infrastructure.

That is why he likes building here. AI gets the good looking chart out fast, and the workbook keeps the security, governance and permissions a corporate setting needs.

The common problems with building a viz extension in Claude are handled in a skill Eric uses for all his extension work. He keeps it global, a viz extension markdown file Claude reads whenever he opens an extension project.

It does not catch everything. This session still hits errors the skill did not cover, but he says his first live extension build had far more in the way.

For the heat map and the Nightingale rose, Eric pastes a screenshot of the chart he wants and tells Claude to recreate it for the flight data, referencing the data file by its copied path. He has also had success starting from a rough mockup.

He runs all three builds in parallel. When he later asked Claude what a full written prompt to recreate these extensions would take, the answer was a lot, which is why he tests screenshots first.

The live builds run on Sonnet 5 at low effort, partly to test the minimum model that still does a good job. For anything complex built from scratch, Eric suggests using Fable for the plan, then switching to a lower model to execute it.

Once the plan is in place, he finds other models run it fine. With Fable he finds low effort works well: faster and far fewer tokens.

The .trex file is an XML manifest that points Tableau at the files the extension needs. During development it points at localhost. To use the extension in a workbook on Cloud or Server, you host the web files somewhere reachable, even inside a corporate network, and change the manifest to point there.

On the Exchange, a partner listing runs on Cloud, Server and locally, but only certified extensions with the badge run on Tableau Public. He names Tableau, LiveDataViz and InfoTopics as the trusted options he has seen.

On the Nightingale rose the marks card reads value, hour and detail, not color, size, text and tooltip. Viz extensions define their own encodings, so Claude built ones that fit the purpose it was prompted for.

Field choices are also set by the code. The heat map only accepts a single hour value, not a datetime, because that is how it was built. You can have Claude accept either.

If the .trex has not changed, clicking the already loaded extension refreshes the web side with Claude's latest changes. If the manifest changed, for example a new marks card, you have to add the extension again, since there is no way to break the cache on the manifest. Fields already dropped on the marks card carry over.

When the beeswarm throws a load error, Eric copies the whole message back to Claude. He says these errors usually mean two extensions are colliding somewhere, and this one turns out to reuse standard card names. It still fails live, so he loads the morning's build.

Viz extensions focus on a single custom chart with its own config. Dashboard extensions, the original type, sit on the dashboard as an object and can read data from every chart on it.

Eric is building a universal search with one: type a field name like carrier to get its values, or type a value like AA and it finds which field holds it, in place of ten dropdown filters.