Julian Neagu

The part about testing your own assumptions really hit home. A green suite can just mean you built a perfect loop between the code and your fixtures. Using real corpus data feels like the better sanity check, especially for security tools where normal words can look suspicious.

MCP makes this much more practical. The AI doesn’t need to know where the data lives, it just needs good tools to ask for it.

The Network tab check is probably the fastest way to confirm this. Seeing a new document request on every tap tells you pretty quickly what’s going on.

I like the idea of failing instead of guessing. A missing citation is easy to spot. A wrong one can sit there for weeks before anyone notices.

I like that you shipped the simple version first. One useful summary comment beats a long feature list that never gets finished.

Health checks are one of those things people leave until later. Then the first outage happens and suddenly they become the most useful page in the whole app.

AI can generate the first version, but knowing what breaks in production is still a human skill. I’ve seen this a lot while building tools. The hard part is making the right calls.

I still like learning sockets by building small things like chat apps. It makes the flow click faster than just reading about packets and protocols.

This is a problem I’ve noticed with long-running agents too. More context helps, but without a clear budget, the agent carries more noise around.

I’ve seen this happen a lot with coding agents. The fix is usually not a better prompt, it’s moving that lesson into the right place so the next session starts smarter.