Don't Lock Yourself to an LLM or Agent Harness

Don’t lock yourself to an LLM or agent harness like Claude, Codex, Cursor, etc. This is plain stupid. I say it again: stupid!

LLMs change all the time. With a new release, your loved prompts and skills might silently expire. In most cases they even make results worse all of a sudden. And it takes some time to find out. These things are not durable at all. And by the way: one-off prompts are a thing of the past if you haven’t noticed yet.

The best way to cope with that is to delete them, start over, and create simple, small and durable skills. Read up on this on Every to learn more.

It’s even more true for the harnesses. They also change. They will change features and system prompts—that’s a huge risk—overnight without you noticing. Your workflows fall apart immediately.

All those suggestions, workflows, skill and prompt libraries out there have an unknown expiration date. This expiration date usually matches new harness or LLM versions. And they change more often than you will notice.

Don’t blindly adopt these “curated” packages. Get the good stuff out of them and roll your own. That’s how you do it. Don’t be lazy.

Most of these Claude, Codex, etc. “operating systems” are plain BS. Put in the work. Don’t be lazy.

What is your experience here?

Originally posted on LinkedIn.