Daily shorts
340 AI shorts in a year, 95.1 million measured views
- lifetime views: 340 shorts + 7 long videos
- 95.1M
- shorts past one million
- 17
- median views per short
- 131k
- best short
- 3.8M
- The problem
- Releasing a 10 to 25-second vertical video almost daily that reads in one second and holds past the first two.
- For whom
- LaRefVlog’s Shorts audience (parody fiction); internally, a test bench that feeds The Oracle and Radar.
- The outcome
- 340 shorts in a year, 95.1M lifetime views with 7 long videos, a 131k median and 17 shorts past one million.
Context
One idea, one punchline, one image readable in a second. Short form is unforgiving: the thumbnail decides the click, the first two seconds decide the rest.
It is my full-scale testing ground. I try a device in front of a real audience before using it elsewhere, and its data feeds The Oracle and Radar.
What I built
- 340 shorts released between October 2025 and October 2026, 195 of them past 100,000 views.
- A fixed format: vertical, Full HD, 20 seconds max, one punchline.
- Python tooling: timestamped transcription, cut detection, contact sheets, measured voice continuity (0.999 correlation).
- Measurement loop: real CTR, daily views and forecasts, fed back into topic choice.
Architecture 6 blocks
- 01 · InputWatchTopics of the day cross-checked across sources
- 02 · ToolScriptOne situation, one punchline, 20 seconds max
- 03 · ModelGenerationFull HD vertical shots, several takes each
- 04 · HumanEditPacing, subtitles, sound; title and thumbnail
- 05 · OutputReleaseOne short a day or so, labelled as parody
- 06 · DataMeasureViews, CTR and retention fed back into the watch
Pipeline 6 steps
- WatchTopics of the day cross-checked across sources.
- ScriptOne situation, one punchline, 20 seconds.
- GenerationVertical shots, several takes each.
- EditPacing, subtitles, sound.
- ReleaseTitle and thumbnail built for the click.
- MeasureCTR, retention and views flow back into the watch.
Decisions the choice, the reason, the alternative ruled out
- D1Fix a written format: vertical 9:16, Full HD, 20 s max, one punchline.
A fixed frame stops piling in more action than 20 seconds can hold and keeps the daily cadence sustainable.
Ruled outReinventing the format for each idea.
- D2Humour built on action and reactions rather than wordplay.
Visible, successive actions read instantly and generate more reliably; fragile visual gags are avoided.
Ruled outHeavily written lines, which hold up poorly in 20 seconds.
- D3Measure voice continuity between videos instead of tuning it by ear.
0.999 correlation between two takes: the voice stays identical from one video to the next.
Ruled outTuning by ear, which cannot be checked.
- D4Keep YouTube Studio data private and show only the models built on it.
CTR and retention belong to the channel owner: they serve decisions, not display.
Results
Views by month of release (shorts and videos)
| Month | Releases | Total views |
|---|---|---|
| Oct 2025 | 68 | 17.1M |
| Nov 2025 | 44 | 12.1M |
| Dec 2025 | 40 | 16.4M |
| Jan 2026 | 35 | 17.1M |
| Feb 2026 | 31 | 8.0M |
| Mar 2026 | 31 | 7.9M |
| Apr – Jul 2026 | 19 | 1.7M |
| Aug 2026 | 29 | 6.7M |
| Sep 2026 | 31 | 4.9M |
Lifetime views as of 18 Oct 2026, grouped by release month (P10 store, dates ±1 day). Publishing paused April to July.
Deliverables 4
- 340 shorts released from October 2025 to October 2026, 195 of them past 100,000 views.
- A written style reference: format, comedic devices, character continuity.
- Python tooling: timestamped transcription (faster-whisper), cut detection, contact sheets, voice measurement.
- Views by release month, drawn from the Radar store.
Gallery 4
What this project means for you
The same know-how, on your problem.
Short-form at cadence
A fixed format, a held cadence and a measurement loop: I can set up the same system for your account, with your topics and your voice.
Test before you invest
Short form as a test bench: validate a character, a device or a hook with a real audience before turning it into a film or a campaign.
Post-production tooling
Transcription, cut detection, contact sheets and voice measurements as local scripts: the repetitive editing steps are automated.