AI-Assisted Development in 2026: Tools, Models and Honest Expectations

Thanks, Adam. The series is finished on the site now, so I can answer with hindsight rather than a guess.

The hardest to hand over were the two parts where the model's first design was confidently wrong. In the staff chat part it put an authorized SignalR hub behind a server-side component, which can never carry the login cookie, so it failed with a 401 before it did anything. In the hardening part I asked it to attack its own app and it found an open self-registration page that it had shipped itself a few parts earlier.

In the booking flow, what I kept was the rules rather than the code. The model wrote the slots, the form and the confirmation codes well. What it never proposed was the thing that makes double booking impossible: a partial unique index, so a cancelled slot can be booked again but a live one cannot be taken twice. That requirement had to come from me, and the race test that proves it had to be written from the spec, not from the model's code.

Parts 4, 6, 8 and 11 on coder000.com have the details, mistakes included.