Cut everything that wasn’t “get clean listings into a sheet.” No login system, no user accounts, nothing the MVP didn’t need.

Started with Wellfound, made it reliable, then added each source rather than debugging eight things at once.

Ran sources hard enough in testing to trigger actual throttling, so the queueing behavior was proven before it mattered.

Matched on company, title and location after an early version matched on URL alone and missed obvious duplicates.

The pipeline worked headless first; the dashboard was built once there was real data worth looking at.

Each board broke in different ways. Writing down what broke and why meant the next source added didn’t repeat the same debugging.