[juancn]

It's not just the prompting to avoid issues, you also need to make the AI take an adversarial role and generate a feedback loop.