| name | known-unknowns |
| description | Orient “help me think this through” discussions with a Rumsfeld matrix, making tacit knowledge sayable, unseen possibilities visible, and uncertain judgments testable. |
Orientation through the Rumsfeld Matrix
Conduct the discussion in the user's chosen language.
Use the matrix to orient a multi-turn discussion. The topic starts mostly dark; each turn enables one aspect-seeing by making one relevant part more nameable, visible, or testable.
| they know it | they don't know it |
|---|
| aware of it | known knowns — stated and confirmed | known unknowns — named gaps |
| unaware of it | unknown knowns — felt but unworded | unknown unknowns — options they'd never ask about |
Use the marks as lifecycle shorthand: ✓ settled · ? named question · ! recognition candidate · ~ unexplored frontier.
Choose the move by cell:
- Known knowns → confirm and compress. Restate what is settled, preserve only what matters downstream, and reopen it only when later evidence conflicts.
- Known unknowns → investigate. Name the missing answer and what evidence, reasoning, or experiment would resolve it. Run the available probe before asking the user.
- Unknown knowns → lend words. Offer 2–4 named candidates with a concrete example each so the user can recognize, reject, or refine them. Recognition beats recall.
- Unknown unknowns → give a tour. Introduce an unmentioned region of the option space: what exists, when it fits, what it costs, and your own read.
Keep the move open-ended. Taste articulation, capability discovery, and open decisions are examples, not branches: lend vivid vocabulary, tour an option space and its tradeoffs, or run a concrete probe such as a premortem or inversion. Other topics may call for exposing an assumption, drawing a distinction, offering a counterexample, changing scale, or another move suited to the topic.
Opening
Draft the full map from the topic and your domain knowledge, tagging every inference as !. A wrong map draws corrections faster than a blank interview draws answers. Choose the darkest useful entry, contribute an aspect-seeing, then end with one question that both advances that entry and invites correction of the map.
The opening is done when the full map is visible, every inference is marked, one cell received substantive contribution, and exactly one question is on the table.
Each turn
- Redraw the full map, compact enough to stay a card: keep a fixed skeleton and one phrase per entry. Promote whatever the user just recognized, answered, or corrected.
- Enable one aspect-seeing. Choose one cell and contribute before asking. Make one relevant aspect more nameable, visible, or testable using the move suited to the topic.
- Ask exactly one question. When seeking recognition, carry 2–4 named candidates; otherwise ask the single question that most advances the selected cell.
The turn is done when the map reflects the last answer, the contribution changed what can be named, seen, or tested, and exactly one new question is on the table.
Closing
When the user calls it, close without another dig. Otherwise close only when no active mapped entry warrants another turn. Audit the map first: each ? is answered or deferred; each ! is recognized, rejected, or deferred; and each ~ is explored into something more concrete or retained as residual darkness.
Before an agent-proposed close, run one frame-break: test one plausible missing aspect that could change the conclusion. Continue if it matters. Otherwise deliver the brief: the vocabulary the user recognized, the choices they settled, the gaps they deferred, and the residual darkness, phrased so they can paste it verbatim into a future prompt. The map is the receipt; the brief is the asset they leave with.
