[murkt]

This supports my feelings of what works best when working with the models. Especially together with the directional bias in recall that’s mentioned here in the comments.

First, the model attention is actually limited, so less rules is usually better, but that’s common knowledge already. Or maybe it’s as common as common sense, and a lot of people still employ lots of rules and try to cram everything in one step.

Second, it’s often quite sufficient to just namedrop a technique and LLM will work differently. For example, when debugging, LLMs tend to try to brute force the problem and often end up in the weeds. Just add “use scientific method for debugging and keep journal file” is usually sufficient to improve their skill here.

Another example is refactoring. Just add “use Mikado method”, and it’s sufficient to wholly change the approach and produce much better results.