Why Your AI Instructions Always Miss the Mark
Deep thoughts on AI and aspirations —— ByteDance Deep Thinking Circle
A common frustration: you write detailed instructions for an AI, but what it produces is nothing like what you wanted. Your first reaction is probably “the AI isn’t good enough,” so you switch models, change tools, rewrite instructions.
But the problem usually isn’t the AI. The problem is: you think you were clear, but you weren’t.
The Map You Think You Have vs. The Actual Terrain
Consider this analogy: what you give the AI is a map; what the AI has to navigate is the actual terrain.
You have a complete picture in your head: what the goal is, what style you want, what’s good and what’s bad, where the boundaries are. But what you actually say is only a small part of that picture. The rest—the parts you didn’t say—the AI doesn’t know. It can only guess. Getting it right is luck; getting it wrong is inevitable.
So “detailed instructions” doesn’t equal “complete instructions.” Writing a lot doesn’t mean you’ve said everything.
Three Types of “Things You Don’t Know You Didn’t Say”
The things you didn’t say fall into three categories, each more hidden than the last.
First, things you don’t know you need to say. The task itself is complex, and there’s information you didn’t even realize you needed to mention. You never thought of it, so naturally you didn’t write it. This is the most dangerous type.
Second, things you think are unimportant. You think “does this even need to be said?” and skip it. But to the AI, there are no “defaults”—if you don’t say it, it assumes it doesn’t exist. What you consider common sense is a blind spot for the AI.
Third, things you want to express but can’t articulate clearly. You have it in your mind, but you can’t put it into words clearly, so saying it is like not saying it at all. You know this is happening, but the AI can’t receive it.
Add these three together, and you have the gap between “the map you gave” and “the actual terrain.” The bigger the gap, the more the AI veers off course.
The Most Practical Solution: Let AI Find Your Blind Spots
How do you close this gap? The most counterintuitive move is: don’t rush into giving instructions—first let the AI help you find what you’re missing.
Ask the AI what you need to clarify: have it list “what else do I need to know to do this task well?” It will lay out your blind spots—the things you didn’t know to say, the things you thought were unimportant. You fill in each one, then let it start.
This step seems slow but actually saves time. Because if you don’t clarify upfront, what gets produced will almost certainly need rework. The cost of three rounds of rework far exceeds the cost of asking once upfront.
Putting It Into Action
Next time before giving AI instructions, do two things:
First, ask the AI what you need to clarify. Use its questions as a mirror to reveal your blind spots.
Second, write out all your “assumptions” one by one—goals, style, boundaries, what’s good and what’s bad. What you think goes without saying is precisely what needs to be said most.
When AI goes off track, nine times out of ten it’s not an AI problem—it’s an incomplete map. Drawing a complete map is far more useful than finding a better AI.
Key points: Detailed instructions aren’t complete instructions; there’s a gap between the map and the actual terrain; three types of blind spots (don’t know to say / think it’s unimportant / can’t articulate); let AI find your blind spots before starting work.