Last reviewed:
What is the map-and-territory principle in AI?
The map is not the territory. What you give an AI (prompt, files, context) is only a representation of the work. The real work happens elsewhere: in the code, the file, the constraints of the ground. The gap between the two is your unknowns. That is where, and almost only where, a mission derails.
The quality of an AI-assisted mission is capped by your ability to clarify its unknowns, not by the power of the model.
01Origin
The phrase belongs to Alfred Korzybski's general semantics: he presented it in 1931 and took it up in Science and Sanity (1933). A map can be useful, detailed, carefully made, and still be wrong about the ground it describes. A subway map is not the subway.
Engineer Thariq Shihipar (Claude Code team, Anthropic) applied it to working with AI in "A Field Guide to Fable: Finding Your Unknowns", published on X on July 3, 2026. The article is about Claude Fable 5; the method applies to current models (Fable 5.1, Opus 5.5, Sonnet 5.5, Haiku 4.5). His observation: the more capable a model becomes, the more the limiting factor shifts from the model to the human. An agent doesn't act on the territory, it acts on the map. Wherever the map is silent, wrong or out of date, the agent decides on its best guess of your intent: often plausibly, rarely rightly.
02The four unknowns
Before launching a task, place it in these four boxes. Each calls for a different strategy. The grid comes from Thariq Shihipar's article; its logic echoes the Johari window (Joseph Luft and Harrington Ingham, 1955).
| What you know | What you don't know | |
|---|---|---|
| Aware | Known knownswhat you write in the request. The foundation, rarely enough on its own. | Known unknownswhat you know you haven't decided yet. |
| Unaware | Unknown knownsthe implicit: obvious to you, never written, recognised the moment you see it. | Unknown unknownswhat you don't know you don't know. This is where missions derail in silence. |
03Why it matters
With a capable model, getting a result is no longer the problem. Getting the right result is. A model produces convincing output even when it misread your intent, which makes the error costly: you find out late.
There are two ways to use AI. Ask it for an answer to copy. Or ask it to help you see what you couldn't. The first weakens your autonomy over time; the second strengthens it. This is the position WYP.agency takes.
04The method, in three phases
The techniques come from the article. Two WYP.agency additions, restating before acting and the candid review, are labelled as such.
Before: frame and reveal
Blind-spot pass to hunt the unknown unknowns, brainstorm and prototype, reverse interview (one question at a time), concrete references rather than descriptions, a risk-sorted plan with your decisions up top. Then execution in a fresh conversation, with the plan attached.
During: hold the course
Implementation notes: when an unexpected case comes up, Claude picks the cautious option, logs the deviation and moves on. WYP addition: have Claude restate the goal before any execution and, to stay in control, ask for a written assumption before each step.
After: verify and harden
An annotated report followed by a quiz, then a pitch document that brings together the prototype, the spec and the notes to win reviewers' approval. WYP addition: an explicitly candid review.
In Claude, since September 16, 2026, the former Chat and Cowork are a single conversation: Claude decides whether to answer or take on the task (gradual rollout, Pro and Max plans first). A setting in the message box controls how independently it works. In Manual mode, the default, it asks before taking actions; in Auto mode, it keeps going, with automated safety checks before each action. Stay in Manual until the plan is approved, and for any action that is hard to undo: sending, publishing, deleting.
A long task almost always fails for one of two reasons: not enough time spent clarifying the unknowns, or a plan with no margin to adjust when the agent hits one.
FAQ
What is the map-and-territory principle applied to AI?
The map is what you give the AI: prompt, files, context. The territory is where the work actually happens: the codebase, the client file, the concrete constraints. The gap between the two is your unknowns, and that is where missions derail.
What are the four unknowns?
Known knowns (what you write), known unknowns (what you know you don't know), unknown knowns (the implicit, obvious but unwritten) and unknown unknowns (what you don't know you don't know). Each box calls for a different strategy.
Why isn't a good prompt enough?
With a very capable model, the bottleneck is no longer the model but your ability to clarify your unknowns. A prompt that is clear about the wrong assumptions produces a plausible but wrong result.
Does this only concern code?
No. Any mission delegated to an AI has its unknowns: audit, writing, strategy, translation. The grid applies wherever a gap exists between the instruction and the ground.
Do you need to be an expert in the subject to use it well?
No. The method is precisely about surfacing what you don't know. You can deliver in a domain you barely master if you reveal your unknowns instead of ignoring them.
How do you reveal your unknowns with AI?
Before executing: blind-spot pass, reverse interview, concrete references, a risk-sorted plan. During: implementation notes, and restating before acting (WYP addition). After: a quiz, a pitch document, and a candid review (WYP addition).
Get the guide as a PDF
The full method guide « map and territory »: 18 pages, the detailed step-by-step and a library of ready-to-copy prompts.
PDF · 18 pages · October 2026 edition · FR / EN / ES · delivered in under a minute
See also
Further reading
Thariq Shihipar, "A Field Guide to Fable: Finding Your Unknowns", X, July 3, 2026
Sources
- Thariq Shihipar, "A Field Guide to Fable: Finding Your Unknowns", Claude Code team, Anthropic, X, July 3, 2026. https://x.com/trq212/status/2073100352921215386
- Know your unknowns, the author's example prompts (GitHub Pages). https://thariqs.github.io/html-effectiveness/unknowns/
- Claude Cowork and chat are now one Claude, Anthropic, September 16, 2026. https://claude.com/blog/cowork-is-now-claude
- Claude Cowork and chat are one Claude, Claude Help Center (Manual and Auto modes). https://support.claude.com/en/articles/16761823-claude-cowork-and-chat-are-one-claude
- Models overview, Claude documentation (current models). https://platform.claude.com/docs/en/about-claude/models/overview
- Map-territory relation (Korzybski, 1931 and 1933), Wikipedia. https://en.wikipedia.org/wiki/Map%E2%80%93territory_relation
- Johari window (Luft and Ingham, 1955), Wikipedia. https://en.wikipedia.org/wiki/Johari_window