Multi-Source Transcript Harvester

A transcript raw-data layer that collects founder/entrepreneur story content with a single command. It takes a sources-manifest.yml + a list of source slugs + a per-source top-N parameter, and routes each source according to its type. For YouTube channels it first resolves the channel ID, lists the uploaded videos, filters by duration (discarding short clips), then fetches the subtitles of the selected videos and converts them into normalized Markdown transcripts. Podcast RSS sources are marked as Phase 4 and silently skipped.

When to use: When you want to feed a founder-pattern library similar to Starter Story / Indie Hackers / My First Million, set up a multi-channel transcript pool for a research agent, or batch-convert competitor/sector content into text. It is not a standalone agent; it is invoked by upstream agents such as founder-pattern-roadmap-advisor or by batch wrappers.

Output: For each video, a normalized .md transcript file with frontmatter, a per-source _harvest-log.jsonl audit record, and a summary envelope (sources processed, number of transcripts fetched/skipped, estimated cost). It is idempotent with resume: on a second invocation only new videos are fetched. With dry_run you can first see and approve the plan and cost estimate. It supports both English and Turkish sources.