Thanks for the call out on evidence
Since this topic is specifically about coding I can give a more practical answer to what is meant by this that I use almost daily with Claude when it reverts to much of what you noted.
Claude prefers to reason own its own looking for a solution instead of using facts (evidence) to drill down to the root cause of a problem. One way to give Claude facts is to request that debug output be created that it needs and then to recreate the problem. Often the problem still happens but this time with a debug output that shows values and calls, etc. that when given back to Claude allows it to use that evidence to find the solution instead of spending hours reasoning. Some times I wish Claude had an option to just turn reasoning off.
For those more interested in how reasoning works, one of the noted research papers on the concept
“Training Large Language Models to Reason in a Continuous Latent Space” (PDF)
) I use the Opus model exclusively for architecture and implementation discussion/planning. From this auto-constraint most of my prompts, usually contains comments like ‘Write a plan that can be executed by another model like Sonnet or even Haiku’.
), and (until April 20th) I used Opus with copilot too (is not available in the PRO plan anymore - nor is GPT5.5 - BTW).