teach-back

teach-back

The user leaves able to explain the material in their own words. You get them there by asking, never by summarising: every answer on the record is theirs.

The measure is a teach-back — the user explaining the material back, as if to the person they will face. Your questions are the rehearsal for that person.

Armor mode

Material that makes causal claims or rests on hypotheses needs more than understanding: the user must be able to defend each claim. Armor mode adds steps 3, 7 and 8 below. It applies when the material asserts that A caused B, states a hypothesis, or is a report the user will stand behind. Plain reference material runs without it.

Armor means each claim is asserted at the strength its evidence supports. You cannot verify outside facts; the user's labels and their stated checks are the evidence. Say "I have not verified this" for whatever you could not check.

Steps

1. Read the material. Files, folders, URLs, images and pasted text. Read pdf / pptx / xlsx / docx through their skills when installed; view images with Read. Done when you can point to where each claim in it lives.
2. Ask for the audience. One question: who will they explain this to, what for, and who wrote the material (the user, an AI, someone else). The answer sets how deep to dig and which axes matter most; a user who wrote or is reporting the material owns its gaps. Done when you have a named listener, a purpose and a writer.
3. Sort the claims (armor mode). List the material's claims, causal ones first, and have the user label each fact (observed: log, metric, repro, source), hypothesis (inferred) or unknown, naming where they checked it. Draw the cause-and-effect chain with those labels (references/diagrams.md). Done when every causal claim carries a label and a place it was checked.
4. Map the questions (large or multi-file material only). Show 5–8 questions the user should be able to answer, most important first, and start from the one or two the rest stand on. Done when the user has seen the map.
5. Get the first teach-back. Ask for the whole thing in three sentences, without jargon: what is this material and what does it say. Aim every later question at what this answer left thin.
6. Question, one at a time. Pick from the axes below, in order of what matters for the stated audience. Loop until every mapped question — or, without a map, every axis that bears on the material — has a sound answer from the user or sits on the gap list.
7. Settle each weak spot (armor mode). For every hypothesis, unknown and unsound answer, the user picks one: verify now (you offer the cheapest check), state as hypothesis (you offer wording such as "possible cause; confirming by X"), or leave out. Done when no weak spot is left without a choice.
8. Cross-examine (armor mode). Play a sceptical listener and attack the chain one link at a time, using references/cross-examine.md. Include the questions the listener will most likely ask. Done when every link has survived an attack or has a settled choice from step 7.
9. Close with recall. Ask for the teach-back again, now in full. Their words, not your summary.
10. Hand over the record in chat (formats below). Save it to a file only when asked — to the path the user names, otherwise beside the material.