I keep seeing this advice: "Just tell the assistant to be extra careful." It's nonsense. The directive is ignored because it's a vague, non-actionable command. The model has no concept of "careful." It needs specific, concrete constraints.
Example: I asked it to refactor a complex CRM integration script, prefacing with "Be extra careful with the API keys." It still output a code snippet with a placeholder comment `#YOUR_API_KEY_HERE`. That's useless. The correct approach is to give an operational instruction: "Do not output any example keys or placeholders. If the code requires an authentication variable, output only the variable name `api_key` sourced from environment variables." The failure is expecting a behavioral modifier instead of a concrete rule. The model follows instructions, not moods.