Anthropic engineer Thariq wrote up what he learned working with the new model Fable, and the core is an old line: the map is not the territory. The map is the instructions, context, and rules you give the AI; the territory is where the task actually happens, the code, the real world, all the constraints you couldn't foresee. The gap between them is where AI falls into pits, goes in circles, and guesses. I connect this framework to the tacit knowledge distillation I do every day: drawing the map clearly first amounts to laying out the judgments you can't quite put into words as explicit rules, and once AI follows them, the success rate on real tasks goes way up.
- People who already use AI every day but often think, "Why did it guess wrong again and make me explain it all over?"
- Consultants, coaches, and knowledge workers who want to hand their professional judgment to AI, only to find the hardest part is "spelling it out"
- Anyone who wants to understand "why the same AI works wonders for some people and keeps stalling for others"
- A grasp of the "map vs. territory" metaphor, and where AI errors really come from
- A map of the "four kinds of unknowns," so you can spot which unknown is your tacit knowledge
- A questioning method you can run before you start, to draw the map clearly, with prompts you can paste straight in
In one line: the map is not the territory
The line comes from an old idea of the linguist Alfred Korzybski. Thariq applied it to working with AI, and I think he put it well.
Here's how he breaks it down. The map is one representation of the thing you want done, that is, the prompts, rules, and context you give the AI. The territory is where the work actually happens, the real setting with all its constraints. What you hand over is always a map, but what the AI has to walk is the territory.
Why AI makes mistakes and goes in circles
If the map is drawn too simply, the AI follows it and, the moment it hits the real terrain, runs into a pile of situations you never wrote down. Now it has only two options: get stuck and ask you, or guess on its own using "standard industry practice." And standard industry practice isn't necessarily the way you want it done.
Thariq puts it precisely: directing AI is a balancing act. Too specific, and it clings to your instructions even when the right move is to turn; too vague, and it fills in whatever it assumes is best. If you haven't thought the unknowns through, you fail on both ends. You don't know which stretch of road has potholes, and you don't know which stretch is actually smooth and you just wish it would turn there.
You just toss out "make me a dashboard." All the AI can do is fill in a pile of assumptions: which fields, who it's for, what style. Get those wrong, and you go back and forth ten times and it's still not quite right.
You spell it out first: who it's for, what questions it needs to answer, which numbers matter most, and where you haven't decided yet. The AI runs off that map and nails it on the first pass; the rest is just fine-tuning.
Four kinds of unknowns: which one is your tacit knowledge?
Thariq splits "what you want AI to do" into four types. The four-quadrant framing itself comes from Donald Rumsfeld, but applied to working with AI, it neatly exposes every spot where "you thought you'd explained it clearly, but hadn't."
What you write into the prompt. You know it clearly and can articulate it; this is the map itself that you hand the AI.
You haven't thought it through yet, but you know you haven't. At least you know to go ask or look it up.
Something so obvious you'd never bother writing it down, yet you recognize it the moment you see it. You know it; you just didn't realize it needed saying. This quadrant is tacit knowledge.
Things you've never even considered. You don't know which piece of knowledge you're missing, and you don't know how good the result could be at its best.
Drawing the map clearly is distilling tacit knowledge
This is the piece I want to add. Thariq is talking about how to align the map with AI, and that act of "drawing the map clearly" is, in the work I do every day, tacit knowledge distillation.
This is often exactly what I do for consultants, coaches, and instructors. They carry a whole set of judgments in their heads, how to open a new engagement, how to spot what matters in a document, where a student's problem is really stuck, and when you ask them "how did you decide that?", they can't quite answer. That set of judgments has been there all along; it's just so familiar it has turned into intuition, and they can't put it into words.
That unspoken judgment is tacit knowledge. Lay it out one rule at a time as explicit rules, and the AI can do things your way. So "drawing the map" and "distilling tacit knowledge" are really two names for the same act. In doing it, you're turning the most valuable, hardest-to-articulate part of your mind into an asset you can reuse, hand to others, and build on.
How to do it: draw the map clearly before you start
In his original piece, Thariq lays out a full set of methods for mining unknowns with AI. I've picked a few that carry over best to non-engineering settings and adapted them into versions for knowledge workers. You don't have to use them all every time; treat it as a toolbox.
Before taking on a new field or new client, just ask the AI to list "the things about this I might not know I don't know," and give it your starting point: what you know and what you don't.
Have the AI ask you one question at a time, targeting the points that would change how you'd do the work. As you answer, you realize how many judgments you thought you'd spelled out were never actually written down. This is the most direct way to distill tacit knowledge.
Some things you simply can't put into words. When that happens, hand over an example you like and let the AI study "how it pulled that off," far more effective than guessing at it with a pile of adjectives.
Once it feels close enough, ask the AI to draft a plan or first cut, putting the decisions most likely to need changing right up front. One look, and you often spot something you hadn't spelled out.
Prompts you can paste straight in
Models keep getting stronger; the map is on you
Thariq closes with a line I strongly agree with: the better the model, the more you can achieve when you use it right. When a long task ends up wrong, it usually means you need to spend more time defining your unknowns up front.
Every blind-spot scan, every interview, every reference is a very cheap way to figure out what you didn't know before it gets expensive to fix. How strong the AI is is one thing; whether you can draw the map well is where the real gap opens up.
So on your next task, don't rush to set the AI to work. Spend ten minutes first asking it to help you surface your unknowns and draw the map clearly. Those ten minutes are you distilling your own most valuable knowledge.