Feed a lecture to your Claude

01Watching a talk doesn’t stick. Auto-subs don’t either.

I watch a talk, I nod along, and six weeks later I can recall maybe two sentences from that hour.

Auto-subs don't fix it. Slides carry charts, before/after redesigns, table layouts, things the speaker just points at and says “this.” A transcript reproduces the blind spot it was supposed to close.

And even a clean transcript fails a generic ask. An agent asked to “summarize the talk” hands me back generic highlights. What I want is deltas: where the talk contradicts, motivates, or sharpens what I already believe and have already written. That means reading my own skill files before I read the source. Summarizers don't do that by default.

So the pipeline splits in two. Mechanical: scripts, no model, run it and leave. Judgment: the reading. The mechanical part exists only to enable the judgment part.

02The mechanical half: transcript × slides

Four steps, one command in kaibot, all local:

Everything downstream reads the bundle, where each sentence sits next to the slide the audience saw when it was said. The video, word JSON, and extracted frames are archive only.

Eyeball the frames before trusting them, no exceptions. Sample four to six spread across the video. Most frames just the speaker's face? Threshold too low, or there's no deck at all. Long gaps through obviously slide-driven sections? Threshold too high. Adjust and re-run. Costs about a minute.

03The notes pass: read the destination first

This is the slow stage: 30+ minutes of close reading per hour of talk. Two rules shape it.

Read the whole target skill before you open the bundle. Read it through one lens: where does the skill err, miss something, or claim without justification? Skip this and you get generic highlights. Do it and every part of the talk sorts onto a concrete gap or into out-of-scope. That sorting is the point.

Read the frame before its phrases, in sequence, taking notes during the read. Meaning builds up over a talk, and later examples depend on earlier ones. Append notes as you go. Do not wait until the end. You cannot hold an hour of material in memory until then.

Extraction is organized into six buckets, and every entry is timestamped:

The grounding rules are hard rules. No claim appears without a timestamp. Numbers and names get re-checked against the frame at the moment of writing, not recalled from a prior viewing. When the transcript and the slide disagree, log a transcript-vs-slide conflict. Do not quietly settle it.

04The digest: one pass, provenance on every rule

Pull the skill-delta sections from every talk in the batch into one working list. When two talks state the same principle, that is stronger provenance, not two edits: one entry, two sources. Give each delta one of four verdicts: add a rule, sharpen an existing rule (usually by adding the speaker's reasoning or a worked example), spin up a new reference file when the material is too big to live in a rule entry, or reject it with the reason recorded; “true but out of scope” is a perfectly good rejection reason.

Two disciplines make the result trustworthy. Draft the whole edit list before you touch the skill. A plan can be vetoed up front; a diff can only be un-picked afterwards. Tag every new or changed rule with its source. Speaker, venue, timestamp: that is how you tell later whether a rule is folklore or has a citation.

Re-ingesting the same topic runs in diff mode. Cheap readers compare each new talk's bundle against the current skill and hand back only net-new material. Expect heavy overlap on a repeat, ~90% already covered. That’s diff mode doing its job. The remaining 10% is the real new material.

End with one file copy: the bundle and notes go into my corpus pipeline's intake, where they are chunked, extracted and made searchable for every future session. Per-slide sections are already shaped correctly for chunking. Skill edits serve the immediate next task. The corpus serves the questions I cannot yet anticipate.

05Steal this: the prompt for your Claude

You don’t need my exact scripts; the pattern needs yt-dlp, ffmpeg, and any word-level transcriber. Paste this to your Claude:

I want to ingest a talk video into (a) a specific skill/doc of ours
and (b) durable notes. Build and run this pipeline:

MECHANICAL HALF (scripts, no judgment):
1. yt-dlp the video at <=1080p with metadata. Keep auto-subs as
   fallback only.
2. Transcribe locally with word-level timestamps (whisper or similar;
   chunk long audio). Pack into phrase lines: [hh:mm:ss] text.
3. Detect slide transitions by sampling at 1 FPS (slow dissolves never
   register at native fps), then capture each frame ~1.5s AFTER the
   transition so the slide is settled. One JPG per slide + an index.
   Show me 4-6 sample frames; we tune the threshold before trusting it.
4. Build bundle.md: one section per slide, frame embedded, every
   transcript phrase that started on that slide filed under it.

JUDGMENT HALF (you, not a cheap pass):
5. Read our target skill/doc END TO END first. Then read bundle.md in
   order, viewing each frame before its phrases. Extract, with a
   timestamp on every entry: principles WITH the speaker's reasoning,
   re-tellable worked examples, verbatim quotes, deltas against our
   named rules, translation notes where their medium differs from
   ours (marked as inference), and a parking lot for out-of-scope.
6. Show me the edit list (rule -> add/sharpen/reject -> one line)
   before touching the skill. Every applied rule ends with a source
   tag: speaker, venue, timestamp.
7. Never write a number or name from memory of the slide; re-open the
   frame at the moment of writing it down.

06For the Claude that just got pointed here

If a human gave you this URL plus a talk they want ingested, here’s your tl;dr.

07Receipts

• Built and scoped 2026-07-05 against two Edward Tufte talks feeding my data-viz skill, then seven more Tufte talks the same week in diff mode. Most of each diff talk was already covered. A thin band of chart-rule deltas went into the skill, and the speculative bits went to a suggestions note instead of padding it.
• The 1fps detection rule exists because the first scoping run at native fps missed most of a 47-minute talk’s slide transitions (slow dissolves never cleared the threshold). The settled-capture offset exists because mid-dissolve frames look like double exposures.
• The chunked-transcription rule exists because an hour of audio in one pass asked the GPU for a buffer roughly double its ceiling and died.
• Every rule in the target skill that came from a talk carries a source tag inline (speaker, talk, year), so a future session can re-check the source instead of trusting my summary of it.

The hooks-and-gates conventions referenced across this series are covered in the companion page, Physics, not pleas.