Working tool · v0.2
Pod Educator
Comprehension-on-demand for podcasts. Pulls captions for a chosen window of a YouTube podcast, opens an AI chat grounded in that window, and durably captures the resulting insights to a personal cognition graph.
What this tool is
Long-form podcasts contain useful thinking, but extracting it on demand is hard. You hear a five-minute passage that matters, lose the surrounding context, and either re-listen (slow) or take notes that immediately decay (lossy). Pod Educator collapses that gap. You paste a YouTube URL and a time range, the tool fetches captions for the window and caches them locally, and a chat session opens with the transcript loaded as grounding context. You ask whatever the moment provoked — the model answers from inside the conversation you were just hearing.
The moment is the question
The defining design choice is that the question is implicit in the act of opening the tool. You don't first have to formulate the question; the time range is the question. The session opens with an automatic overview of the window, so the conversation starts with shared context rather than empty input.
Insight capture
A pinned button on every chat message saves that exchange as a durably-stored insight in a personal cognition graph. The save includes AI-suggested metadata — title, tags, signal-strength classification, decision implications, and links to related prior insights — that the user can accept or edit before commit. Each session's insights enter a structured representation subsequent sessions can recall from; the graph accumulates a searchable, surface-able, citable record of the operator's thinking over time.
Why this is a lab artifact
Pod Educator is a working consumer of the cognition-graph infrastructure the lab develops. It is small, it is built, and it is in regular use. Each session typically produces a handful of saved insights; the cumulative graph is itself a working example of what structured insight capture looks like when treated as first-class infrastructure rather than a notes folder.
Current state
v0.2. Local-only Streamlit application; runs at localhost:8501. YouTube
caption fetch via yt-dlp; transcript cache survives upstream caption
changes. Stable enough for daily use; not yet packaged for external distribution.
Reading the work
For substantive questions, or interest in adapting the architecture to a different source surface (long videos, lectures, recorded meetings) — tim@cruxadjacent.com.