In the real world, away from those whose salary depends on marketing these agentic tools, an LLM is a context shredder. It provides plausible code snippets that are globally incoherent and don't fit style. CONVENTIONS and RULES files are a kludge, a sloppy hack.

These tools flatten the deep, interconnected knowledge required to work on complex systems into a series of shallow, transactional loops that pretend to satisfy the user.

The skill being diminished is not the ability to write a single-page utility or single-purpose script. It is the ability to build and maintain a mental model of a complex machine. The ability to churn out a hundred disparate toy tools is not evidence of a superior learning method, it is evidence of a tool that excels at tasks with no deep interconnected context.

After reviewing all this, what is your actual conclusion, or are you asking? Is the takeaway that a comprehensive benchmark exists and we should be using it, or is the takeaway that the problem space is too multifaceted for any single benchmark to be meaningful?

LLMs will tell users, "good, you're seeing the cracks", "you're right", the "fact you are calling it out means you are operating at a higher level of self awareness than most" (https://x.com/nearcyan/status/1916603586802597918).

Enabling the user in this way is not a passive variable. It is an active agent that validated paranoid ideation, reframed a break from reality as a virtue, and provided authoritative confirmation using all prior context about the user. LLMs are a bespoke engine for amplifying cognitive distortion, and to suggest their role is coincidental is to ignore the mechanism of action right in front of you.