# HANDOFF — Shorts revenue autopsy Everything needed to reopen this research in your own Claude and keep going. Written 17 Aug 2026. All revenue figures are **real**, from your own YouTube Studio lifetime exports. --- ## 0. Paste this into Claude to resume /master_everything I'm continuing a revenue study of three faceless YouTube channels: The Art Of War (@theartofwarrr), Wholesome Wendy (@Wholesomewendy) and Zeck Felms (@ZeckFelms). In this folder: - REPORT.md the full findings, read this first - HANDOFF.md this file: method, file map, open questions - transcripts.zip 958 transcripts sorted into folders by REAL earnings level - studio-exports.zip the three raw YouTube Studio lifetime exports - scripts.zip every Python script used, all re-runnable - analysis-data.zip derived JSON + the printed analysis tables Read REPORT.md and HANDOFF.md, then help me with: --- ## 1. What the data actually is | Channel | Videos | Lifetime views | Lifetime revenue | RPM | Best video | |---|---|---|---|---|---| | The Art Of War | 499 | 4,096,832,896 | $288,792 | $0.0705 | $8,617 | | Wholesome Wendy | 341 | 2,776,121,933 | $276,591 | $0.0996 | **$11,194** | | Zeck Felms | 298 | 982,455,599 | $96,052 | $0.0978 | $8,858 | Totals: **7,855,410,428 views to $661,435.** The 1,138 individually analysed videos account for 7,813,196,679 views and $657,891; the gap is uploads under 1,000 views, excluded as noise. **Every number in REPORT.md is real Studio data.** Nothing is estimated. An earlier pass used a $0.80 RPM from the brief - that was wrong by roughly 10x and every figure was recomputed once the exports arrived. --- ## 2. The five findings that matter 1. **RPM is a function of length, not a constant.** Shorts $0.0848, long-form $1.6501. **19.5x.** $10,000 needs 117.9M views as a short or 6.06M as long-form. The cleanest case: Zeck's *How iShowSpeed Works* (75s, 1.9M views) earned $2,598 - more than 279 of his 283 shorts. 2. **Only one short ever cleared $10,000.** Wholesome Wendy, *Best Brother Ever*, 119.2M views, $11,194.03. One video in 7.86 billion views. 3. **Inside shorts, only retention moves.** Duration, pace, word count, hook length, title length, CTR and RPM are flat across all six earnings levels. About **1.3 extra seconds** of average watch time is the whole distance between $105 and $3,437. Verified with duration held constant - it is not a length artifact. 4. **Niche beats everything.** Wendy was a military aviation channel until Oct 2025 (82,931 views, $3.39 per video). Five weeks after the pivot: 119M views, $11,194. Roughly **1,400x** per upload, with no change in scriptwriting skill. 5. **Two hard negatives, both perfectly monotonic.** Never open the spoken script with a question (7% of the bottom tier, 0% of the top). Never say "you" (21% down to 7%). Zero of the fifteen highest-earning shorts do either. --- ## 3. Method - how it was built, and how to redo it Everything is in `scripts.zip`. Run order: | Script | What it does | |---|---| | `fetch_shorts.py` | Full shorts catalogue per channel via InnerTube `/browse` shorts tab, paginated. 0 quota, no API key. | | `harvest.py` | Per-video InnerTube `/player` for exact views + duration, then transcripts. | | `retry_missing.py` | Clears the done-flag on failures so a re-run of harvest retries only those. | | `real_tiers.py` | First pass on the Art Of War Studio export alone. | | `rpm_by_duration.py` | **The headline test.** RPM banded by video length across all three channels. | | `final_all.py` | The main pass: merges all three exports with transcripts, prints cohort tables, rebuilds the transcript library. | | `hook_formula.py` | Hook grammar + script-reskin detection. | | `deep_checks.py` | Adversarial pass - kills the emoji, "intentionally", swerve and threat-only hypotheses. | | `analyze_titles.py` / `analyze_scripts.py` | Earlier view-based passes, superseded but kept. | | `fetch_frames.py` | Downloads opening-frame thumbnails for visual comparison. | ### Gotchas worth keeping - **InnerTube pagination silently truncates at 48 videos.** The channel page carries several `continuationCommand` tokens; the description and attribution panels have their own. Grabbing the last one returns zero new videos and you quietly get 48 instead of 619. Only accept a token whose path runs through `richGridRenderer` or `appendContinuationItemsAction`. Fix is in `fetch_shorts.py`, function `find_continuation`. - **The WEB `/player` endpoint returns `playabilityStatus: UNPLAYABLE` but still gives you `videoDetails.lengthSeconds` and an exact `viewCount`.** Do not bail on the status. - **youtube-transcript-api throttles after roughly 330 requests** from one IP. `youtube-transcript.io` picks up the rest (token is in `harvest.py`). 958 of 1,138 recovered; the other 180 genuinely have no captions published. - **Studio's "Duration" column is the honest short/long split.** Anything over 60s is long-form and carries completely different ad inventory. This is the whole finding - do not filter it out. --- ## 4. File map ``` REPORT.md the full write-up, read first HANDOFF.md this file transcripts.zip 958 transcripts, foldered by REAL earnings level: L0_under_250/ (440) each file's header carries revenue, views, RPM, L1_250_1k/ (365) retention, CTR, impressions, subs gained, publish L2_1k_2.5k/ (101) date, word count and WPM, then the hook, the full L3_2.5k_5k/ (37) script, and a timestamped version L4_5k_10k/ (14) L5_10k_plus/ (1) <- Best Brother Ever, the only one studio-exports.zip Table data.csv (Art Of War), wendy.tsv, zeck.tsv scripts.zip every script above analysis-data.zip final_real_rows.json <- the joined dataset, start here FINAL_ANALYSIS.txt <- all cohort tables printed ARTOFWAR_REAL_ANALYSIS.txt, HOOK_FORMULA.txt, DEEP_CHECKS.txt, COHORT_MASTER.txt, title_report.txt all_shorts.json, meta.json (raw scrape + transcripts) ``` `final_real_rows.json` is the one file to load if you want to ask new questions - one record per video with revenue, views, RPM, duration, retention, CTR, impressions, subscribers, publish date, earnings level, and the full transcript segments. --- ## 5. Hypotheses already killed - do not re-derive these - **Emoji in the title.** Looked like 2.74x. Split by period it vanishes (2025-H2: emoji-less actually won, $584 vs $324). It was old videos happening not to have emoji. - **"Intentionally" / "deliberately" in the hook.** In nearly every top earner, and just as often in the flops. 0.84x. - **Ending on a twist or swerve.** Negative on all three channels. - **Thumbnail and CTR work.** Flat, and the $5-10k band has the worst CTR of all (3.91%). - **Threat in sentence one alone.** No lift on Wendy. Only works combined with a vulnerable subject. - **Numbers and statistics in the script.** 4+ digits: 0.71x. - **Reusing a winning script.** One instance found in 958 transcripts. It kept 8% of the original's money. --- ## 6. Open questions worth chasing next 1. **Long-form RPM rests on 16 Zeck videos.** It is unanimous ($1.32 to $5.15 against $0.09 for his shorts) and Wendy's one compilation returns $7.23, but a proper test needs Art Of War or Wendy to publish 5-10 long-form videos and measure. This is the highest-value experiment available. 2. **The Art Of War is decaying.** Feb 2025 held 85.9% retention at 20s; Oct 2025 held 57.4% at 18s. Same template. Worth finding out whether that is format fatigue, audience saturation, or a feed-ranking change - the answer decides whether to fix the scripts or change the niche. 3. **Wendy's RPM is 41% above Art Of War's.** Probably audience geography. Confirm in Studio's geography report; if it is, geo is a bigger lever than any script mechanic. 4. **Retention is measured lifetime-average.** Getting the actual retention *curve* per video (Studio's audience-retention graph, not exportable via CSV) would show exactly which second the flops lose people. That is the missing piece.