Major flaw in current LLMs is them forgetting clear and simple instructions.

This is not about the context window. At least, it really does not seem like it.

You tell it to fix all tests, "do not stop until 100% completed". It agrees, fixes a few, and 2 minutes later congratulates itself for reaching 96.9% 🎉

In context files, we still need to use over the top language like **CRITICAL** or it just doesn't give a fuck.

Looking forward to the next-gen of "don't make me repeat myself" LLM tech.