Articulate · video programme
YouTube — objective, research, plan
Reframe
Not reach. A video that can be defended.
- Brief was "a video that gets 1,000 views". Now: a video Anthony can be proud of and defend.
- Views = audience-size problem. Defensibility = craft problem.
- Ten channels at 0.3–5.8M subs cannot teach the first. All cleared 1,000 views years ago.
- They can teach the second — craft is visible in the artefact.
- Corpus is now a standards study, not a growth study.
Objective · Research · Plan
One
Objective & goal
- One video that survives scrutiny. Every claim sourced on screen.
- Test: would Anthony defend it in a client room.
- Register: authoritative, evidenced, argued. ColdFusion / Nerdrotic lane.
- Length: 21–24 min.
- Subject: Saudi Arabia 2026, from assets already gathered.
- Real deliverable: a repeatable method Articulate can sell.
- Not the goal: subs, virality, RPM, a channel business.
Two
Research — can get
- Full back-catalogue — IDs, titles, durations, dates
- Views, likes, comments → engagement rates
- Most-viewed ranking within a channel
- Thumbnails at max res, bulk
- Transcripts with cue timings
- Script metrics — pace, evidence, argument, narrative, address, variety, hook load
- Asset metrics — face/screen split, cut rate, PiP geometry
- On-screen text via OCR
- Comment corpora — API keyed 27 Aug
- Growth — forward from today only, self-snapshotted
Three
Plan
- 1 · Standards pass. 3 videos/channel, not 10.
- 2 · Asset layer at n≈25. Where the information still is — script metrics have converged.
- 3 · Teardown of the Saudi cut against the bands.
- 4 · Rebuild the script to hit them — in Anthony's argument, not a borrowed one.
- 5 · Ship, then measure against the same rubric. Fails its own test → doesn't go out.
- Rule: every step runs local or free by default.
Cannot know
Behind auth
- Retention curves, avg view duration
- CTR, impressions
- Traffic sources — browse / search / suggested
- RPM, CPM, revenue
- Needs owner-scope Analytics API
- Social Blade + vidIQ estimate revenue. ViewStats has no API.
Recorded nowhere
- Historic sub growth for channels we don't own
- Why a video won — no counterfactual
- What was cut in the edit
- Whether the thumbnail was A/B tested
- Honest proxy: public most-replayed heatmap
The real gap
- The 0 → 1,000 mechanism. Not in this corpus. All ten passed it years ago, at lower production values.
- Whether craft transfers from film/geopolitics to professional services. 8 of 10 are entertainment.
- We measure what they made, never why it spread.
The bands
- Argument density: 12.2–14.8 connectives per 1,000 words. Six independent channels. ~1 reasoning move every 70–80 words.
- Holds across economics, technology, film criticism, military affairs, business.
- Duration: 14–26 min, with a hole where 9–13 min would be.
- Both confirmed twice, by independent selection methods.
- These are the standards to build against.
Corpus
- Ten channels, selected from 1,858 watch events in Anthony's browser history — not recall.
- Filtered to median upload inside the long-form argument band.
| Channel | Subs | Med. dur | Med. views | Role |
|---|---|---|---|---|
| Preston Stewart | 1.09M | 14m | 169,500 | Most-watched · zero narrative · high address |
| Nerdrotic | 1.29M | 19m | 537,500 | The register at scale |
| Nerdrotic Daily | 304K | 19m | 67,500 | Same voice, volume cadence |
| Disparu | 348K | 23m | 126,000 | Dead centre of the band |
| Moon | 1.76M | 22m | 411,000 | Best views-per-sub |
| Critical Drinker After Hours | 511K | 18m | 89,000 | Conversational variant |
| Rick Beato | 5.81M | 23m | 766,500 | Authority case — expertise shown |
| Joe Blogs | 467K | 16m | 39,000 | Economics explainer |
| Emergency Awesome | 4.78M | 18m | 275,500 | High-cadence speculation |
| Project Kamp | 749K | 26m | 137,000 | Control — long-form, not argument |
Skill — video-teardown
- Any YouTube URL or channel in → measured structure out.
- Derived metrics only. No transcript text stored. Clean on copyright.
- Cost £0.
| Layer | Measures | Route | Status |
|---|---|---|---|
| Script | wpm · evidence/1k · argument/1k · narrative/1k · address/1k · variety · hook load | yt-dlp → yt-structure.py | Built · run on 6 |
| Asset | face/screen split · cut rate · PiP geometry · text load · source-shown rate | ffmpeg sample → classifier → OCR | 3 built · OCR pending |
| Packaging | title pattern · thumbnail composition · face presence · text load | Data API + thumbnail fetch | Spec'd, not built |
Compute
| Work | Runs on | Why | Cost |
|---|---|---|---|
| Harvest, metadata, thumbnails | Sandbox scripts | Pure I/O — no model at all | £0 |
| Rubric metrics | Sandbox regex | Deterministic, built | £0 |
| Re-punctuation | Local · Studio | High volume, low judgement, checkable | £0 |
| Beat-map extraction | Local · Studio | Structured extraction, fixed schema | £0 |
| Synthesis, teardown, argument | Frontier | Carries the judgement — not delegated | Metered |
- Scope cut: 3 videos/channel, not 10. Thirty transcripts, not a hundred.
- Bands were converging at n=6. A hundred buys precision on a number we won't act on differently.