Another thing I noticed is the llms perform efficiently/effectively only under the optimum circumstances. May be this just a Claude issue, but when the session goes on for very long the effectiveness drops drastically and I start getting bad answers. This is especially true for design and debugging. Wonder how that ladders up to the token usage