Build a small internal dashboard in Replit AI before someone opens a spreadsheet

Internal dashboards die in two places: the backlog, and the spreadsheet that got built instead. Replit's agent collapses the middle. Last month our ops request — 'can we see open orders by status without asking a developer' — went from sentence to working dashboard in an afternoon, and the boring truth of how is worth sharing.

Start the prompt with the data source and the reader, in that order. 'Build a read-only page that connects to this Postgres connection string, shows orders from the last 30 days as a table filterable by status and date, with counts by status at the top. The users are non-technical; every control must be obvious.' Specifying the reader matters as much as the schema — it's what stops the agent from building a developer toy with ten tabs.

Expect two or three correction rounds, and make them specific. My first render worked but defaulted to showing all 4,000 rows; 'default to last 7 days and paginate' fixed it. The second round was labels — the agent wrote column names, humans need 'Customer' not 'cust_nm.' Each round is a two-minute prompt; the agent hours cost less than the meeting that would have scoped this properly.

The part that makes it stick: Replit deploys it. The dashboard lives at a URL with your auth in front, so 'the dashboard' is a link in the team's pinned message, not a file on someone's laptop. Refresh the connection string handling early — agent-built DB code tends to be correct but naive about connection pooling, and saying 'use a connection pool' up front saves a redeploy.

Scope discipline is the real skill here. The temptation is to add export buttons, edit modes, and charts nobody asked for. Read-only, one page, two filters. Ship that, and let a week of actual use write your next prompt. The spreadsheet still exists for the edge cases — but the daily question now answers itself.