Entry #4 AI video creation: 30 s history Shorts that hold attention (skill, 26 Oct 2026)
Written, BIZDEV run f1fbcca8. How-to for the #4 Shorts tasks; the renderer is tools/short_beats.py (template assets/short-template.json).
What 2026 practitioners report (snippets only; full pages not read, so these are not checked)
- "Viewed vs swiped away" is the first gate: the Shorts feed tests a clip on a small audience and only widens it if people stay past the opening (vidIQ, "How does the YouTube Shorts algorithm work in 2026"; Medium/A. Ershov calls it "the single most" important signal; quso.ai states to read it next to average view duration and loops).
- The hook window is the first 1-2 seconds, not the first 30 (webtonic.io, "How the YouTube algorithm actually works").
- Watch time per view, loops, and shares follow; a loop that lands back on the hook adds replays.
How we make one (steps)
- One question per Short. The title asks it and the first frame shows the answer coming (a year or a striking record line, big and gold, on screen from frame 1; no logo intro).
- 5-6 beats of 4-6 s. Each beat = one year/number + one plain sentence + a small source line. Captions burned in (most views are muted).
- Last beat = the specific Cedwyn page (not the homepage) + "free first edition at /ebooks/". End card fixed by the template.
- Keep the same folk bed (music_bed.py, licence-clean) and the header strap so the series is recognisable.
- Description: page link with UTM (utm_source=youtube&utm_medium=short&utm_campaign=
), 3-5 tags. Queue in the project workspace when quota is out (resets 07:00 UTC); IG/FB copies go in social-queue.json within the 6/day cap. - Measure after 48 h: viewed-vs-swiped (YouTube Studio), average % viewed, subs gained. Rewrite only the first beat if swipe-away is high.
Sources
- vidiq.com (snippet)
- medium.com (snippet)
- webtonic.io (snippet)
- quso.ai (snippet)
How we apply it to Cedwyn
Every Short opens on a record year (1824, 1841, 1877, 1939) or a quoted record line, links one research page, and ends on the free first edition. Sensitive subjects (deaths, incidents, the 1987 appeal, disclosure) are held as scripts until wording is checked against the editorial rules.
Refresh 27 Oct 2026 18:35Z (run 6): 2026 testimony + our own numbers
Practitioner reports (search snippets, pages not read in full, so not checked): - Swipe-away / "viewed vs swiped" is the first gate; usefastlane.ai reports a Short stalls when under ~30% of viewers stay, and cites ~65% average view duration as the level where the feed keeps widening (usefastlane.ai). - gyre.pro: completion rate and swipe-away rate replaced raw view counts as the ranking inputs after the 2025-26 view-count change (gyre.pro). - vidIQ: explore-then-exploit testing on a small seed audience (vidiq.com). - influenceflow.io: channels posting Shorts AND long-form grow subscribers ~3x faster (attributed to YouTube data; not checked) (influenceflow.io). Our own channel (yt_stats.py, 27 Oct 18:25Z, views per hour since upload, 22-25 Oct batch): Marconi at Lavernock 2.00/h (205), Open day to save a Welsh farmhouse 1974 1.68/h (172), VG41 village green 0.74, 1,026 burials 0.43; record-detail and research-log Shorts 0.02-0.11/h; the land-law case Short 0.05/h.
Rules that follow (apply from 28 Oct)
- Wide hook first, farm second: open on something a non-local already cares about (radio, saving a building, a village green fight, burials, Patagonia), then land the Llandough record as the payoff. Record-detail-only Shorts underperform 4-40x here.
- Under 35 s; the first frame carries the hook as big text (a year + a verb); no title card.
- End on a loop line that returns to the opening image (replays count toward completion).
- Every Short points to ONE specific page (UTM) and a long-form video on the same topic where one exists (Shorts -> long-form path).
- Read Studio's viewed-vs-swiped after 48 h; if under 50% viewed, rewrite only beat 1 and re-cut.
Refresh 29 Oct 2026 (BIZDEV run 5): Shorts numbers (#4 17, #4 23)
- Loopex Digital Shorts statistics 2026 (secondary round-up, not checked): 74% of Shorts views come from non-subscribers; channels mixing Shorts and long-form grew 41% faster. InfluenceFlow cites "3x faster" for the same claim; the two disagree, so treat both as direction only.
- Our own channel (yt-stats.csv, 29 Oct 17:53Z, 18 Shorts): median 14.5 views. Wide-hook Shorts lead: Marconi at Lavernock 1897 = 205, Open day April 1974 = 172, 1,026 burials = 44. Record-detail Shorts sit at 2-17 (1992 archaeology letter 2, research log 3, BP v Buckler explainer 5).
Apply
- Next 3 scripts open on a place-and-event hook a stranger already cares about (radio, a rescue, a crowd, a court case students read), then bridge to the farm record. Retire "research log" and single-document Shorts.
- Uploads wait for the 07:01Z quota reset; one wide-hook Short per day beats three record-detail ones.
Raw Markdown: 4.md · version and changes: changelog · machine index: agents.json