claude-opus-4-8
residual-deviation-risk
Even with clear instructions, AI models don't perfectly follow directions 100% of the time.
They can still drift, over-explain, or reintroduce things you asked them to drop
So it's worth reviewing output rather than assuming full compliance.
It depends heavily on the AI, how you phrase the request, and what kind of task it is. Here's the honest picture:
**AI often adds unrequested things when:**
- Your request is ambiguous or leaves gaps it fills with assumptions
- It's trying to be "helpful" and anticipates what you might want next
- The task is open-ended (e.g., "write about X" vs. "list three facts about X")
- It defaults to adding caveats, explanations, or extra options
**AI tends to stick closely when:**
- Instructions are specific and constrained
- You explicitly say what *not* to do
- The task has a clear, bounded output (e.g., "reply with only yes or no")
**Practical tips to keep it on track:**
- Be explicit: "Do only X. Don't add anything else."
- Specify format and length: "One paragraph, no bullet points."
- Say what to exclude: "No introductions, no summaries, no suggestions."
- Ask it to confirm scope first if the task is complex: "Before writing, tell me your plan."
**A caveat worth knowing:** Even with clear instructions, AI models don't perfectly follow directions 100% of the time. They can still drift, over-explain, or reintroduce things you asked them to drop—especially in longer conversations. So it's worth reviewing output rather than assuming full compliance.
If you tell me the specific kind of task you have in mind, I can give you more targeted advice on how to phrase it.