Every rubbish internal tool starts life the same way. Someone opens a design app, drags some pretty boxes around, builds a dashboard, and only then, right at the end, works out where the data is meant to live. That order is why the thing falls over the moment real work hits it.
Angus Sewell (Angus the Nontechnical) makes the fix his whole argument: build data-first and treat the screen as the last, throwaway layer. His nine-word version is hard to beat. “The screen is ephemeral. The data is the product.” He walks through a four-step build, design the data, wire Claude to it, prototype the screen, then host, and it is a solid recipe. One nugget worth stealing from him: the official Postgres connector for Claude is read-only and the repo was archived in 2025, so if you want write access you need a community one running in unrestricted mode. That single gotcha saves an afternoon.
He is right about the order, and I want to add one step in front of his, because data-first can still fail if you skip it. The dashboard research is blunt about why these projects die, and it is not the schema. It is that nobody asked what decision the tool is supposed to support before they started building. You can design a beautiful, normalised database and still build something useless if it answers questions nobody needed answered.
So my version is decision-first, then data-first, then screen-last. Write one sentence: this tool exists so that a named person can decide a specific thing faster. Everything downstream, the tables, the columns, the constraints, falls out of that sentence. Design the data around the decisions and the query patterns, not just around the nouns in your head. Skip the sentence and you are simply building a tidier version of the wrong thing.
The deeper reason this matters in 2026 is the one Angus is circling. A clean data model does more than please humans. It is the surface an AI agent can actually operate. A screen is for a person to click. A well-structured table is something you can point Claude at and say do the thing, and it will. The prettier your dashboard and the messier your data, the less of your work you can ever hand off.
Draw the screen last, and draw it in pencil. It is the cheapest, most disposable part of the whole build, and treating it as the main event is exactly why your tools let you down.
Did you know: schema-first design, where the data structure is the single source of truth rather than an afterthought, has been shown to improve stability, privacy controls and data lineage, the unglamorous stuff that decides whether a tool survives contact with real users.