How can I use system instructions or a CLAUDE.md file to prevent AI from crossing the line into giving advice or making determinations?

One approach discussed is to explicitly instruct the AI in system instructions not to give advice—stating clearly that it should not provide any kind of advice. From experience, AI models comply with this about 95% of the time, which at least demonstrates a good-faith attempt. One member described using a CLAUDE.md file that tells the AI who they are, how they operate, and what considerations to take into account—framing it as "you are meant to help me do my job, not do my job for me." They also maintain a markdown file of vetted regulatory source material that the AI references as source of truth, with instructions to avoid actions that might cross compliance lines. The broader point made was that being set up properly—having detailed, granular instructions about firm-specific processes—improves both output quality and the ability to operate cautiously. However, it was noted that in this profession, "close enough" isn't good enough; practitioners need to be certain, not just approximately compliant.