The bitter lesson is being played out in coding agents: don't over-invest in skills as every new generation of coding LLM can solve coding problems previous generations couldn't. Skills/rules/claude.md are priors injected as context for your LLM - they can help solve specific coding problems, but can also reduce general purpose coding abilities. LLMs are like general purpose compute with search/optimization - they get better at solving coding tasks than earlier models with skills.
The coding agent bitter lesson for developers is to review and pare back your skills for every new generation of coding LLM.
https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models