Environments & test data
| Thing | Value |
|---|
| Playbook (read-only surface) | format-cloning-youtube-playbook_2026-09-30 — verified HTTP 200 2026-09-30. Repo source: docs/research/ in microspaces/superstarsale. |
| Video staging | Unlisted uploads are the staging. Every pipeline artifact ships to YouTube as unlisted first; public visibility is only switched after B5/C2 pass. Nothing unpolished ever hits the public channel. |
| YouTube (no staging for public) | A brand channel starts with zero videos and zero subscribers, so the first public upload is the test. One niche per channel — the POV wealth/history ladder is the only format this channel carries. |
| Money rails | AdSense + YPP (Mike-only: payee, tax, 2FA). No affiliate links anywhere on this channel — monetization is AdSense RPM plus nothing in round 1. |
| Pipeline tooling (verified on this machine) | ffmpeg 9.0.1 (C:/Users/mikeb/bin/ffmpeg) for assembly; TTS via the agent; yt_dlp (yta-venv) for competitor research pulls; script/template generation agent-side. Pipeline code lands in the superstarsale repo (or its own repo once it exists — the plan page updates when it does). |
| Pick (from the playbook) | Format: POV narrative ladder — "Your Life at Every Rank/Level of X" (verified active cluster). Niche: money/wealth history and storytelling, $10–15 RPM band, story-not-advice. Second channel (phase 2): sleep/lean-back long-form for watch-hour volume at $3–8 RPM. |
| Policy backdrop (verified 2026-09-30) | AI disclosure toggle mandatory for realistic synthetic voice; July 16, 2026 inauthentic-content policy excludes template-identical uploads and AI personas giving finance/health/legal advice; YPP = 1,000 subs + 4,000 watch-hours/12mo, doubling to 8,000h for new applicants from Feb 1, 2027. |
Test data markers
- Test uploads: unlisted, titled
pairtest-- — e.g. pairtest-2026-10-13-pipeline-smoke. Nothing titled pairtest- ever goes public.
- Topic list: the 30 planned public topics live in the repo (
docs/research/faceless-topic-list.md once created); test topics never enter it.
- Cleanup: every test artifact starts with
pairtest- — see the checklist at the bottom.
The Google account, the channel, and AdSense are real business assets, not throwaways. Nothing in this plan requires deleting them at session end — cleanup targets test uploads, test descriptions, and any stray test comment.
Never apply to YPP below the thresholds: a declined review locks reapplication for ~30 days. With the Feb 1, 2027 bar doubling, a wasted 30-day lockout is the single most expensive mistake this plan guards against (D4).
Flow A · Playbook & decisions — the factory floor plan (no account)
Persona: the operators re-reading the plan before building. Read-only on the live playbook + repo. Produces: confirmed decisions everything downstream depends on.
A1BothOpen
https://microspaces.github.io/superstarsale/research/format-cloning-youtube-playbook_2026-09-30.html fresh (hard reload), desktop.
Expected: HTTP 200, title reads "Faceless YouTube Format-Cloning Playbook — 2026-09-30" (a "Daily YouTube Strategy Research" title means a stale Pages deploy — stop and reconcile), verified-claims table renders with ✅/⚠️ marks, policy section present, watch-hour math section present.
A2BothConfirm the pick is recorded and the 30-topic list exists with
substance-distinct entries: same skeleton ("Your Life at Every Rank/Level of X"), 30 different searchable subjects, no two topics a viewer could confuse.
Expected: format + niche match the playbook's Recommended pick; the topic list exists in the repo and every entry is a distinct real subject (this list is the inauthentic-content defense — B5 gates on it). A list of near-duplicates (e.g. "wealth levels", "wealth levels part 2") is a fail.
A3AgentRe-run the watch-hour/revenue math from the playbook sources: 2.0h @ 50% retention → views per 4,000h; 10-min @ 50% → the same; revenue at $3/$10/$15 RPM per 100K views.
MikeSpot-check one number against the playbook table — if agent math and published math disagree, that's the issue.
Expected: 4,000 views (2h) vs 48,000 views (10-min) — a 12x advantage; $250/$1,000/$1,500 per 100K views at $2.50/$10/$15 RPM. All numbers reproducible from the playbook's formula, no "60x" anywhere.
A4BothDeadline tracker: compute days remaining to Feb 1, 2027, and the required watch-hour run-rate to clear 4,000h before then (uploads/week × expected hours per upload batch).
Expected: both testers get the same dates and a run-rate that is achievable at the planned cadence; the number is written into the thread and becomes D2's tracking baseline. If the run-rate needs more than the planned cadence, the cadence decision changes now, not in December.
Flow B · Pipeline — one topic in, one finished video out (the content machine)
Persona: QA on the factory. Feed data: one topic from A2's list. Produces: script, voiceover, assembled MP4, thumbnail — the exact artifacts Flow C uploads. Repo/pipeline level; nothing touches YouTube yet.
B1AgentGenerate the script for the B-topic from the format template (same skeleton as the verified cluster: hook → the ladder, rung by rung → payoff).
MikeRead one full script cold. Judge it as a viewer: would you keep watching at rung 3?
Expected: script follows the locked skeleton, covers its one subject in depth, reads as storytelling (how wealth works at each rung) — zero advice language ("you should invest…") anywhere. Any advice phrasing is a B5 gate failure, fixed before TTS.
B2AgentRender TTS voiceover for the B1 script; measure runtime and listen-check the first 60s for glitches.
MikeListen to a 60s sample on phone speakers — the lean-back audience is on TVs and phones, not studio monitors.
Expected: audio renders end-to-end, runtime scales with script length (~150 wpm narration), no clipping/truncated tail, sounds natural enough that a cold listener doesn't bail in the first minute.
B3AgentAssemble the full MP4 with ffmpeg: voiceover + visuals + chapters, loudness-normalized, ends cleanly.
MikePlay the file locally; skip around (25%, 50%, 75%) — every point must have picture + sound.
Expected: plays in a plain browser/phone player, runtime matches B2 audio + padding, chapters present, no silent or black stretches at skip points, file size sane for runtime.
B4AgentGenerate the thumbnail for the B-topic per the format's visual convention (real searchable subject, no fabricated faces/logos).
MikeView it at phone-thumbnail size (~10mm wide) — can you read it? Would you click it next to the verified cluster's thumbnails?
Expected: legible at thumbnail size, subject matches the title's searchable entity, style consistent with what a viewer of the format cluster expects. Illegible-at-a-glance = regenerate before C3.
B5AgentPolicy gate (blocks C3): grep the B-artifacts for advice language ("you should", "invest in", "buy this"), confirm the disclosure flag is planned ON, and diff the topic against the rest of A2's list for substance-distinctness.
Expected: zero advice hits in script/description; AI disclosure marked for the upload; topic unique on the list. A failure here is fixed in the pipeline, not waived — this gate exists because the July 2026 inauthentic-content policy demonetizes at channel level, and finance-advice personas are an explicit excluded category.
Flow C · Channel & first publish (Mike-led; agent verifies from outside)
Persona: the business going live. Feed data: Flow B artifacts. Produces: a real channel, a staged test upload, and the first public video — the assets Flows D and E probe. These are Mike-only actions; the agent cannot create accounts or pass Google 2FA.
C1MikeCreate the YouTube
brand channel under the business name (per the playbook identity — handle + name consistent with the format), set channel art and description, turn on 2-step verification for the Google account.
AgentRecord the handle in the thread; verify the channel resolves logged-out (incognito) — public, art and description present.
Expected: channel live and public; name/handle consistent with the pick; brand-channel type (not a personal rename) so ownership can be shared later without handing over Mike's Google login. A private or unfinished-looking channel is a fail — the first public upload lands on it.
C2MikeStaged end-to-end upload: upload the B3 MP4 titled
pairtest--, set
Altered content → Yes (the AI disclosure toggle), visibility
unlisted, paste the description + chapters.
AgentMike shares the unlisted link: fetch it, confirm title marker, description intact, chapters present, disclosure label visible on the watch page, and the video is
not publicly searchable (search the exact title logged-out — zero hits).
Expected: unlisted = link-accessible but unsearchable; disclosure toggle survives the upload; description/chapters intact end-to-end. This proves the full pipeline (B artifacts → live YouTube video) before anything public exists.
C3MikeFirst public upload: topic #1 from A2's list, B-gated artifacts, title per the locked skeleton, disclosure ON, public.
AgentLogged-out: confirm public, title matches the skeleton exactly, thumbnail is B4's (not an auto-frame), disclosure label visible on the watch page, description clean (B5's zero-advice check holds on the
published text).
Expected: the channel's public face is exactly the factory output — no manual re-edits that bypass the pipeline. Disclosure label visible without expanding anything. Any human "quick fix" made at upload time is a process miss — file it.
C4Both48h analytics baseline: Mike reads Studio (views, watch time, average view duration, impressions CTR); agent records the same numbers from Mike's screenshot into the thread + the log.
Expected: numbers exist in Studio (even if tiny) and match Mike's screenshot; average view duration is the number that matters — it feeds D2's run-rate math. A first upload with 0 impressions after 48h is itself a finding (title/thumbnail/metadata issue), not a reason to panic-delete.
Flow D · Watch-hour rail & cadence (the road to monetization)
Persona: the channel as a compounding asset. Feed data: C's live channel. Produces: the lean-back upload pattern, a watch-hour tracker vs the YPP bar, and a deadline-safe application plan.
D1BothLong-form/lean-back upload: a 2h+ video through the same pipeline (B1–B5 on a sleep-compatible variant of the format — calm narration, loopable), staged unlisted (C2 pattern) then public.
Expected: pipeline handles 2h+ without runtime/audio drift; disclosure ON; the video plays past the 1h mark without artifacts. This is the unit that makes the watch-hour math work (4,000 views ≈ 4,000h at 50% retention — 12x better than 10-min content).
D2BothWatch-hour tracker: cumulative public watch-hours vs the 4,000h bar and the A4 run-rate, updated per upload from Studio exports.
AgentCompute from Mike's Studio screenshots; post the tracker to the thread weekly.
Expected: one shared tracker, both testers read the same numbers, trajectory vs the Feb 1, 2027 deadline visible at a glance. Tracker math recomputable from the screenshots (no black box).
D3AgentCadence + format-lock audit over everything published so far: every title fits the one skeleton; topics stay substance-distinct (A2's list, checked off); uploads ship at the planned cadence.
Expected: format is locked (a viewer browsing the channel sees one coherent show), substance is distinct per video (the inauthentic-content defense holds in practice), cadence matches the A4 plan or the plan is formally revised. Skeleton-drift ("let me just try a different style this week") is a fail — clone the format forever, vary the substance.
D4MikeYPP application, executed deadline-safe: check Studio's monetization progress; apply
only when both bars (1,000 subs + 4,000h) are green — and before Feb 1, 2027 so the 4,000h bar applies. Agent tracks the calendar and posts the countdown.
Expected-failure pinned: if the bar is missed and 8,000h applies, record the new math and the revised run-rate as the D2 baseline — the plan absorbs the doubling. What must never happen: applying below threshold (30-day reapplication lockout, the red callout in Environments).
Flow E · Policy & compliance probes (agent-led)
Persona: the platform reviewer / an adversarial audit. Mike observes and re-runs the in-app ones. These steps are expected to pass silently — any hit is a content or pipeline fix before more videos publish.
E1AgentAI-disclosure sweep: every public upload's watch page shows the altered/synthetic content disclosure; the toggle was set at upload, not retrofitted.
MikeRe-check from the phone app — labels can render differently in-app than logged-out web.
Expected: 100% of public uploads disclose. Undisclosed realistic synthetic voice is a YPP-removal risk (labels get applied for you; repeat offenders lose the program). One miss = pull the video, re-upload disclosed, file the process miss.
E2AgentAdvice-language sweep on every published script, description, and pinned comment: no "you should buy/invest/choose", no portfolio or product recommendations, no AI-persona-giving-financial-advice framing.
Expected: zero advice anywhere — the channel tells stories about wealth, it never advises. Finance-advice personas are an explicit excluded category under the July 2026 policy; this is the channel-level kill-switch.
E3AgentSubstance-distinct sweep: pair up the published library — no two videos cover the same subject or a reskinned subject. The skeleton repeats; the substance never does.
Expected: every pair distinguishable by subject matter, not just title wording. This is the inauthentic-content defense in practice — "same script structure re-rendered with different nouns" is the exact excluded pattern, and the topic list (A2) is the proof it can't happen by accident.
E4AgentBorrowed-authority audit: every real name/place/entity used in titles, scripts, and thumbnails is real, searchable, and treated factually — zero fabricated entities, zero false claims about real people.
MikeSpot-check one title's entity: search it yourself — does the real thing exist and match how the video uses it?
Expected: every borrowed hook resolves to a real searchable subject (the playbook's "never fabricate" rule); nothing implies a real person endorsed or appeared in the video. A fabricated entity is a same-day takedown.
E5BothIsolation + session cleanliness: the channel carries zero affiliate/store links (this business monetizes via AdSense only — cross-contamination with the IG rail is a fail in both directions); no
pairtest- upload is public; unlisted tests stayed unlisted; no test comments left anywhere.
Expected: one business per channel, one monetization rail per business, zero test-marker leaks to public surfaces. Any pairtest- strays are listed for cleanup in the checklist below.