Dustin Mihalik uses job search to explain the limitations of placing a conventional web interface inside an AI chat. If an app fetches and displays data without returning that information to the model, follow-up questions about results become difficult to answer. He recommends keeping structured tool content and the rendered interface consistent.
Dustin Mihalik describes updating tool descriptions so the model knows that results are already visible, reducing redundant text output. He also explains why user interactions need to update model context: opening a job description changes what the user is looking at, and the model needs that context to summarize it or help with a related task.
Dustin Mihalik argues that combining every search with immediate rendering can discourage deeper exploration and produce too many result panels. His proposed architecture separates text-based search tools from a rendering tool that accepts selected job IDs. The model can then search across locations, filter candidates and render a smaller relevant set, with tool descriptions specifying where valid rendering inputs originate.
Dustin Mihalik extends this pattern to model-generated explanations and highlighted details within a controlled interface. His closing recommendation is to design small, composable data and rendering tools around what the model needs to explore, allowing the final UI to reflect that work rather than constrain it.
Watch on YouTube
