Lab · 4
Can AI build a helpful filter, sort, and search?
Experiment 3 found that AI can fix drift, but only once someone points it at the specific thing to fix — it doesn’t notice on its own. This experiment tested something more contained: could AI design and build a genuinely useful piece of interface — filter, sort, and search — working inside the site’s own atomic-design rules, and make it run identically across two different pages?
Hypothesis
AI can easily create a helpful tool for sorting, filtering and searching content.
Finding
AI can easily create a helpful tool for sorting, filtering and searching content.
Tested by building and shipping the tool itself, live on /lab/ and /thinking/ — not a controlled study across other sites, other content types, or other design systems.
Method
What was built, and how the work split between AI and Dr. Paul.
What was built
One new organism — a content toolbar, combining search, filter, and sort — assembled entirely from atoms and molecules already documented at the design system: a search field, a filter chip, a sort select, and a result count, grouped into a filter group and a sort control, and finally the toolbar itself. No new colors, radii, spacing values, or type sizes were introduced. The same component, same markup, same script, now runs on both /lab/ and /thinking/ — only the filter facet and sort options change per page: Status for the Lab, Topic for Thinking.
AI handled
- Designing the components inside the atomic-design constraints
- Writing the CSS and the filter/sort/search logic
- Building a working mockup for review before anything shipped
- Wiring the toolbar into both live pages
Dr. Paul handled
- Deciding a shared toolbar was worth building at all
- Choosing the filter facet for each page
- Reviewing and approving the mockup
- Setting the hypothesis, finding, and framing for this page
Finding
AI can easily create a helpful tool for sorting, filtering and searching content.
The component came together in a single session, drawing only on tokens the site already had — no new visual language, and nothing simplified to fit the constraint.
The same organism runs both pages from one shared script; only the filter facet and the sort options differ per page.
Implication
For a CMO, this is a cost story more than a feature story. Findability tooling — filter, sort, search — is usually the work that loses to campaign deadlines for the same engineering hours, so it sits on the backlog until a content library is already hard to navigate. If AI can build it to spec, on brand, and production-ready inside a single session, it stops being a backlog item and becomes a checkbox: every content-heavy page can have one the day it ships, not the quarter someone finally gets to it.
The tools that make content easy to find shouldn’t be the ones AI can’t be trusted to build. This one, it built well.
Next Question · 5
Can AI build an infinite-scroll page that stays fully visible to AI?
This experiment proved a shared component could run identically across two pages. The next one asks something riskier: could AI take on a UX pattern that’s traditionally made content harder for machines to find — and get it right for AI answer engines, not just search?
Read Experiment 5 →