The jobs to turn down in an AI studio
A shop that takes everything does its worst work at its thinnest margin. Disqualifiers across brief type, client sophistication, approvals and exposure.
9 min read
The mouth goes soft after a lip-sync pass
Older lip-sync models rebuild a tiny mouth crop and paste it back soft. Finish with a face restore, or switch models when the patch is structural.
8 min read
The separate limiters behind agent and social calls
Agent chat, generation and social posting sit on separate counters. Size each stage against its own limiter so a publishing 429 cannot stall generation.
9 min read
The opening text card as a second hook
On-screen text in frame one hooks independently of the audio and can quietly contradict the visual. How to test card variants against one identical cut.
9 min read
The order of finishing passes: clean, then scale
Going straight to 4K amplifies artifacts you could have removed cheaply. The pass order that works, and what each pass can and cannot fix.
9 min read
The paperwork file every campaign needs
A per-job records structure for generated work: what to capture at delivery, which three records everyone skips, and a retention clock that starts on use.
9 min read
The perception gap between you and your audience
Ad executives overestimate younger consumers' warmth toward AI ads by 37 points. A per-brand tolerance test you can run in a fortnight.
8 min read
The refresh flag on repeat workflow runs
First runs keep the scenes you approved. Repeat runs rewrite the plot unless you pin them. What refresh regenerates, and what it holds constant.
8 min read
AI Training Fair Use: Ross Has Not Ruled
The Third Circuit has not ruled in Thomson Reuters v. Ross. Ross was non-generative, so the appeal does not settle fair use for generative AI training.
8 min read
The 70/20/10 pillar split, measured
Treat 70/20/10 as a starting prior. Tag videos by audience overlap, measure each pillar, and reallocate the mix when the numbers stop matching the split.
9 min read
The 30-55 second window for clipped content
Repurposed clips cluster between 30 and 55 seconds because shorter loses context and longer loses completion. A trim target and a way to test your band.
9 min read
The Virality Predictor as a Pre-Publish Gate
Chasing an absolute virality score is a dead end. Running the predictor on two cuts of the same video is a gate that decides which one ships.
9 min read
There is no Veo 4: Google's real video lineup
There is no Veo 4. Google's video work runs across Veo 3.1, Gemini Omni Flash, Flow and Google Vids. Here is which of those you can actually name in a brief.
8 min read
Dear Algo puts Threads reach in users' hands
Threads' 2026 Dear Algo and Your Algo controls let people ask for more or less of a content type. How to be the thing they ask for more of.
8 min read
How the Threads feed mixes followed and recommended
Meta describes the Threads feed as an ML-ranked blend of followed and unfollowed accounts. What changes when you write for the half that has no context.
8 min read
Threads' five-minute video limit
Threads takes five-minute video. Shorts caps at three; Reels stops recommending past three. Formats that need the extra minutes, and a padding test.
9 min read
Threads has no creator monetization program
Threads runs no creator payout at all. That absence removes your demonetization exposure, which makes it the cheapest place to test formats before they earn.
8 min read
The EU's Three AI Disclosure Icons
The EU's marking Code of Practice finalised three official disclosure icons. Where they belong on a video, and why they stack with platform labels.
8 min read
Three pairs minimum: beating feed variance
One paired short-form test is mostly feed lottery. Run three or more pairs in the same hook family and read the aggregate, not the lucky clip.
8 min read
Three ways to make a face talk, and when each wins
A native-audio model, a lipsync pass, or a photo-to-avatar render. Compared on realism, script length ceilings, and what each costs when the copy changes.
8 min read
Tier deliverables by difficulty, not duration
Runtime stopped predicting effort. Build a rate card on the nine shot attributes that actually drive attempt counts, with tier bands and an attempt allowance.
9 min read
TikTok's allow high-quality uploads setting
TikTok compresses harder on mobile data and hides a high-quality upload toggle. Find it by name, upload on wifi, and stop re-encoding a clean render.
8 min read
TikTok bans AI likeness of private people
TikTok bans a realistic AI likeness of a private adult without permission, and of anyone under 18 even with it. A label does not make the video allowed.
8 min read
Caption length shrinks TikTok's safe zone
TikTok's own ad docs say the usable frame contracts as caption text grows, which makes caption copy a layout input you decide before the export, not after.
9 min read