Education & TrainingFree

SME Interview to Training Material

The interview, checked against an independent breakdown of the procedure, with a question against every step it never reached.

River takes the procedure the training has to teach, and the interview you recorded about it. It writes its own breakdown of that procedure before reading a word of the transcript, then compares the two. What comes back is a module outline drafted from what the expert actually said. Beside it sits a register with one row per step: stated, supplied by you and agreed to, or absent from the recording entirely. Every row that is not stated carries the question that would close it.

How much gets left out has been measured, once, carefully. Ten trauma surgeons described an emergency shunt procedure, and the descriptions were scored against a protocol an eleventh surgeon approved. The ones working from memory alone omitted nearly 70% of necessary decision steps. Hold the number loosely, because the comparison arm was a single interview and the setting was surgery rather than a compliance module. What it establishes is that the gap can run to most of the procedure.

Built for the instructional designer with a recording and a deadline, and the enablement lead converting a retiring specialist's know-how. It also fits anyone told to make a course out of an hour of someone talking. What ranks for this tells you to ask better questions and record more of the answer, which is sound and is not a coverage check. If the recording is unclear rather than incomplete, a transcript workup fits better. If a whole course landed on you instead, start with what you inherited.

The file says when and what, never who

Absence is not the only thing a transcript hides. Two people were in the room, and the file your recorder exported may not say which of them spoke. Every cue in a WebVTT file must carry a timing and a payload, while the format's one speaker construct, the voice span, is optional. The specification maps that name to a title attribute, which a browser renders as a hover tooltip rather than as text. SRT, which the same specification names as its own basis, has no speaker field at all.

That matters more here than it does almost anywhere else. A step belongs to the expert only if the expert raised it. Ask whether the batch gets re-run on a mismatch, hear yes, and the transcript now holds that step in your words, where a flat read cannot tell the difference. So the run separates what was stated from what was supplied and agreed to. Where the file carries no attribution at all, it says so rather than assigning turns by ear.

Briefing the expert better is the obvious alternative, and the research is discouraging. Experts anchor on their own knowledge when judging what a task costs a beginner. In one study, people learned facts until their own accuracy rose from .43 to .88, while their estimate of a novice's score barely moved, from .53 to .57. The authors also note earlier work finding that warning people about this bias does not remove it. Neither study tested this kind of interview, so treat them as a reason to check coverage rather than a measurement of it.

How it works

Name the procedure
Attach the transcript or caption file, and say what the training has to teach and who is being trained.
Decompose it first
River breaks the procedure into steps and decisions before it reads a word of your transcript.
Compare the two
Every step is matched to where the expert covered it, or marked as never reached in the recording.
Send the gaps back
What is missing leaves as a short list of specific questions, not a document to review.

What you get

One row per step in the procedure: stated by the expert, prompted by you, or absent
The module outline, sequenced for someone learning it rather than for how it was explained
A question against every step that is not stated, short enough to actually get answered
Decisions the expert made without naming a threshold, listed with the threshold left blank
Where the file carries no speaker, the turns it could not attribute, marked rather than guessed
Timecodes on every stated step when the capture format carries them, and absent when it cannot

Common questions

My SME sent 147 slides. How is anything missing?

Both things are true at once. What an expert hands over is what they count as content: history, edge cases, the whole deck. What they skip is what they count as too obvious to say, and that is where the decisions live. Volume and coverage are different measurements, and the register reports the second. A long deck can still miss the number that decides everything.

Is this saying my SME is bad at their job?

The opposite. Skill that has gone automatic is what expertise looks like, and the steps that vanish are the ones performed without deliberation. A review addressed to instructional designers puts it plainly: experts' free recall of strategies introduces errors and omissions into instructional materials that hinder student success. The register is a list of questions nobody thought to ask.

Where does the list of steps to check against come from?

River writes it, which is the honest answer and the limit worth knowing. The breakdown comes from the procedure you named and the audience you described, not from a standard anyone approved. So the rows are candidate gaps and a person accepts or rejects each one. The surgical study scored against a protocol an expert signed off, and this has no equivalent.

My transcript has no speaker labels. Does that break it?

It costs a column, not the run. Coverage still reads out of the text, because a step either appears or it does not. What becomes unreliable is telling a step the expert raised from one you raised and they agreed to. Those rows come back marked unattributed rather than assigned, since guessing which line sounds more expert is how a prompt becomes a fact.

One expert. Can it tell me how common a gap is?

No, and it will not pretend to. A single interview has no frequency in it, so nothing here counts how many people mentioned a thing. Counting across a sample is a separate discipline with its own rules. Every row here is one step and its evidence: where the expert covered it, or that the recording never reached it.

Can it just write the module for me?

It drafts one, and the draft and the register do not carry equal weight. The outline restructures material the expert actually provided, so it rests on evidence. The register claims an absence, inferred against a breakdown River wrote, so it is weaker. Both arrive labelled that way. Sessions and evaluation live in a separate pack, and the SCORM build specification comes after.

Will my SME actually answer the questions?

Nothing here promises that. Experts are busy, and a full write-up sent back for review tends to be skimmed. The register is short for that reason: ten specific questions, each with its step attached, are easier to answer than a document to review. River can turn them into a short email or a fifteen-minute follow-up interview script.

Sources

Built by

Chandler Supple

Co-founder & CEO, River

Chandler has spent the last six years working in AI, building tools for millions of users and Fortune 100 companies. In 2025, he co-founded River to build a new kind of operating system for AI agents.

LinkedIn GitHub X

Reviewed by

Brian Joseff, Co-founder, River

LinkedIn X