Field guide
What goes wrong when you build software with an AI coding agent, and what we do about it. Each page takes one failure we've run into while building our apps with Claude.
It said it was done, and it wasn't
Claude will tell you the work is finished when it isn't. Here is how we make every "done" pass through something that could have said no.
It lost the plot in a long chat
The longer a chat with Claude runs, the less of the start it really remembers. We stopped keeping anything important in the chat.
It keeps making the same mistake
Claude starts every session knowing nothing about your last one. A lesson only sticks if it's written where the next session will actually read it.
It built the wrong thing really well
Claude will build whatever you describe, beautifully, including the thing that can't work. Test the risky idea with a real experiment before any app gets built.
It believed something that wasn't true
Claude reasons beautifully from whatever it's told, including things that are wrong. The cheapest check that could settle a question goes first.
It works but feels off
Claude is good at making things work and much less reliable at making them feel right. Feel needs its own review, and the final say stays with a person.
Running several agents at once made a mess
Several Claude sessions working side by side can do a day's work in an hour, or wreck each other's work in a minute. Each one needs its own space and the warnings that come with it.
It burned through my usage limit
Most of what eats a usage limit is text nobody typed. Measure what Claude is actually reading before you try to cut it.
It did something I didn't want
The real risk is rarely a dramatic rogue AI. It's a quiet step you can't take back, or a tool that did the wrong thing and reported success. Keep the one-way steps for yourself.