optimizing internal prompting for small models like OSS-20B

[COWI-MI]

The long and complex instructions for the models seems to be optimized for the frontier models, not small models like OSS-20B that is drowning / loosing track in complex instructions.

Has anyone had success with changing the instructions and possible other steps to get small models to perform better?

Se example of instruction for OSS-20B:
https://github.com/openai/codex/blob/8cf5b00aef0d4eda507f1b03ae4b669a6a8013d5/codex-rs/core/models.json#L618C1-L618C21106