Maybe the solution is to cull the bad, cliché writing from the training data.
You can just instruct the LLM not to write like an LLM.
alt Hacker News
You can just instruct the LLM not to write like an LLM.