Finding the demonstration inside the podcast
I worked on ChatPRD engineering and the How I AI plugin, including a way to connect written workflows back to the demonstrations in the episodes. Codex searched a pilot set of transcripts and made a queue of moments to inspect in the videos. The useful output was the shortlist, with the evidence beside each candidate.
Find workflow demo candidates in podcast transcripts with Codex
I wanted readers of How I AI to be able to jump to the moment a guest demonstrates a workflow. I had Codex inspect the transcripts in Dropbox and compare the passages with our published workflow catalog.
The pilot produced 23 candidates across four episodes and ten workflows. Each candidate kept its supporting transcript excerpt, a time window, and a reason to review it. That made a much more useful starting point than a list of timestamps without any explanation of what was supposed to happen there.
The transcripts also exposed a problem with treating timestamps as finished links. Some only had minute markers. Others came from a raw recording or an edit with a different opening. A passage could identify the right demonstration while still pointing to the wrong place in the published video.
The result was a review queue, not a set of verified clips. None of those cuts had been checked against video yet. The transcript search narrowed what to inspect; watching the published edit was still the next step before adding a “watch demo” link.
# Find workflow demonstrations in transcripts
Use the supplied episode transcripts and published workflow descriptions. Match episodes by their content and identity, not just folder order or similar titles.
For each workflow, locate passages where someone actually demonstrates the task. Distinguish a demonstration from a passing mention or a story about earlier work. Return the supporting excerpt, transcript time window, what the viewer should see, and confidence in the match.
Inspect the timing format. Minute markers are search windows, not precise cut boundaries. Raw and draft recordings may differ from the published edit. Leave timestamped video links unset until verified.
Rank a small pilot for review. List unread transcripts and workflows with no supported match rather than implying complete coverage.
Where video access is available, confirm the published video identity, compare spoken anchors, inspect the screen action, and refine the boundaries to complete thoughts. Clearly distinguish transcript candidates from video-verified segments. Do not publish or modify the catalog.