Orchestrator Unknowns (Finding Your Unknowns)
Once a model is capable enough, the binding constraint on the quality of its work stops being its raw intelligence and becomes the things the orchestrator has not specified — the unknowns the human providing the instructions, plans, and prompts carries into the task. The source relays an Anthropic field guide ("A Field Guide to Fable: Finding Your Unknowns," attributed on-screen to Thariq / @trq212) whose framing is the map is not the territory: your prompt, skills, and context are the map; the codebase and real-world constraints are the territory; an "unknown" is any place the map underspecifies the territory, forcing the model to guess what you want. The field guide's claimed load-bearing line is that Fable is the first model where the quality of the work is bottlenecked by the operator's ability to clarify its unknowns, not by the model's ability.
The organizing device is a four-box matrix (the Rumsfeld quadrants applied to prompting): known knowns — what is already in your prompt; known unknowns — the gaps you can name; unknown knowns — tacit knowledge obvious to you but never written down (which even a very capable model may lack, since it does not share your life experience); and unknown unknowns — what you never thought to ask. The claim that makes this actionable: your prompt only fills the first box; you can use the capable model itself to drag the other three into the light — eliciting the questions, tacit assumptions, and unexplored territory a plan would silently paper over.
This is adjacent to but distinct from the Imagination Constraint. That concept is about what to build — the ceiling on value is the size of your list of askable things. This concept is one project down: given a task you have chosen, the ceiling on how well the model executes it is the unknowns you failed to clarify. The supply story for the tacit/unknown boxes is Tacit Capability Awareness: you only know where to probe by having hands-on touch with the model's edge.
Claims
- Past a capability threshold, the bottleneck on a model's output quality is the orchestrator's unclarified unknowns, not the model's raw intelligence. principle — durable once models are capable enough: when the model can execute, the residual error concentrates in what the human underspecified, so the leverage moves to specification. The source attributes the sharpest form ("the first model bottlenecked by my ability to clarify its unknowns") to the Anthropic field guide.
- Every unknown falls in one of four boxes — known knowns (in your prompt), known unknowns (gaps you can name), unknown knowns (your unwritten tacit knowledge), unknown unknowns (what you never thought to ask); the prompt only fills the first. observation — the field guide's framing device (Rumsfeld quadrants applied to prompting), reproduced as the source presents it.
- Use the capable model to surface the unknown knowns and unknown unknowns — elicit the tacit assumptions and unexplored territory rather than only answering the questions you already know to ask. best practice — context: working with a model strong enough that your specification, not its ability, is the limiter; the move is to spend its intelligence on eliciting unknowns, which pays off only when the model genuinely exceeds your ability to enumerate the gaps. Pairs with Frontier Scouting.
- "The map is not the territory": your prompt/skills/context are a representation; the model must guess wherever the representation underspecifies the real constraints, and more work means more unknowns hit. principle — durable framing: the gap between specification and reality is where a generator's guesses (and errors) live, independent of model generation.
Related
- Imagination Constraint — adjacent axis. That concept is what to build (the ceiling is your list of askable things); this is how well a chosen task is executed (the ceiling is the unknowns you left unspecified). Both say the human, not the model, is now the limiter — at different altitudes.
- Tacit Capability Awareness — the supply side for the two hidden boxes: you can only sense which unknowns matter (and which tacit knowledge to write down) by having touched the model's capability edge.
- Wargaming (Adversarial Contingency Planning) — the operational answer: a wargame's forks and RECON-NEEDED markers are the mechanism that drags the named and unnamed unknowns into an executable plan.
- Intent Context — supplying the why of a known request; this concept is broader — surfacing the gaps in the request the operator didn't know to state at all.
- Frontier Scouting — the practice of deliberately probing for unknown unknowns rather than only executing the known list.
- Distillate: Do THIS Before You Lose Access to Fable 5 — war-game the missions, keep the blueprints