# AI Prompts Are Maps Not the Territory ## Claim A prompt, skill, or context package I give an AI agent is a representation of the task, never the task itself, so it should be treated as provisional until checked against the real system. <!-- atomicity: 1 idea ✓ --> ## Scope & conditions - Holds for any AI-agent workflow operating on a real, messier-than-described environment: a codebase, a production system, a community, a governance process. - Breaks down as a useful distinction only when the "territory" genuinely is fully specified by the map — narrow, closed-world tasks with no external state to diverge from. - Does not tell you how often to re-check the map against the territory, only that the gap is structural and won't close on its own. ## Evidence Korzybski's "the map is not the territory" applied to agentic AI: prompts, skills, and context are the map; the codebase, production environment, and real constraints are the territory. The map is necessarily narrower and cleaner than the territory, so agent output that looks confident can still be wrong about the parts of the territory the map never captured. > "The map, a representation of the work to be done, is my prompts and skills and context, it's what I give Claude... The territory is where the work needs to happen, the codebase, the real world, its actual constraints." ## Links - [[AI & the Future of Knowledge Work MOC]] — the epistemics of instructing an agent — the specification is never the task - [[3. Resources/Permanent Notes/Watching a System Run Beats Reading It]] — because extends: watching real execution is exactly how you catch where the map diverges from the territory. - [[3. Resources/Permanent Notes/Verification Must Test the Claimed State Not Only the Latest Change]] — because extends: verification against the territory has to check accumulated state, not just the agent's latest claimed diff. - [[3. Resources/Permanent Notes/A Feedback Signal Not More Automation Is What Makes a System Self-Improving]] — because applies-in: the map-territory loop is a specific instance of feedback-driven correction. - [[3. Resources/MOCs/AI Agent Safety MOC]] — because applies-in: treating agent plans as provisional maps is a baseline safety posture, not an optional refinement. ## Tests / falsifiers - Find a real Fable/agent session where the plan (map) matched the eventual working result (territory) with zero correction loop — a genuine case would weaken the claim that a loop is always necessary. - Track whether teams that explicitly budget for "territory-checking" (tests, log review, diff inspection) ship fewer regressions than teams that trust the plan as delivered. ## Maintenance - Last tended: 2026-08-16 - Next review: 2026-11-16 ## Source - [A Field Guide to Fable Finding Your Unknowns [01kx13e2dfjeh2vjxhzr7e19g2]](3.%20Resources/Readwise/Tweets/A%20Field%20Guide%20to%20Fable%20Finding%20Your%20Unknowns%20%5B01kx13e2dfjeh2vjxhzr7e19g2%5D.md)] — states the map/territory distinction as applied to working with Claude Fable 5.