Method

How this was built

One snapshot, eight agents, three rules: every metric traces to a file in data/, no quotes longer than 15 words, no fabricated numbers.

Pipeline

Acquire. yt-dlp pulled per-video .info.json, auto-captions (.vtt), top 200 comments, thumbnails, and channel metadata.
Normalize. A Python pass built data/videos/summary.jsonl — one line per video — with per-cohort view z-scores (long vs short separately), heuristic topic clusters, beat segmentation, and CTA detection.
Analyze. Seven subagents ran in parallel against fixed slices of data/: algorithmic-positioning, content-forensics, hook-retention, seo-packaging, audience-community, brand-identity, niche-competitor.
Sequence. An eighth agent, growth-experiments, read all seven reports and ranked twelve experiments by leverage × ease.
Synthesize. The orchestrator merged everything into the 90-day roadmap on Synthesis.

Hard rules followed

No fabricated numbers. Every metric traces back to a file in data/.
No re-downloading from subagents. If an agent needed more data, it wrote a .requests.md and halted.
Cite video IDs in brackets, e.g. [abc123XYZ].
No verbatim quotes longer than 15 words. Paraphrase aggressively.
Direct claims only — no "fascinating insights," no "in conclusion."
Per-cohort z-scores. Long-form and shorts are scored separately so they're never ranked against each other.

Known limits

60 of 80 videos sampled. YouTube's bot challenge blocked 20 videos mid-acquisition. The 60-video set is treated as authoritative; the 25% sample shortfall is flagged in every report.
53 of 60 transcripts. Above the 80% gate. Three videos with no transcript were excluded from hook-pattern counts.
Peer set is aspirational. The 8 peers in data/competitors/ are 25×–1,800× larger than The Dossier. There are no acquired peers in the ideal 100k–500k "next rung" band.
Peer view counts are unavailable. Top-20 flat playlists from yt-dlp do not carry per-video view counts or upload timestamps. Niche comparisons use title and duration patterns only.

What this is not

Not a live dashboard. It is a single snapshot.
Not an SEO keyword tool. No keyword-volume data was used.
Not a content generator. The roadmap proposes experiments; humans execute them.

Source data acquired via yt-dlp. Channel under analysis: @thedossier_in · Snapshot date: 2026-05-17 · 60 of 80 videos sampled (YouTube bot challenge blocked the remainder). Every metric traces back to a file in data/.