independent AI and markets research founder @contangresearch
Joined August 2025
- Tweets433
- Following216
- Followers52
- Likes4K
4 rules we stick by when a real product goes into a fully illustrated ad:
1. pick the mode before you prompt — photoreal product in a drawn world, or product sculpted too. Mixing the two mid-set is what stops four panels reading as one campaign.
2. the material mismatch is the point. Everything is plasticine, the mask is not. That break is what makes the eye read a real object instead of a prop.
3. the problem owns the frame, the product takes 15-20%. Panel one is a sleeper ripping a mask off at 3AM. The mask sits on the nightstand.
4. never let micro-text on the product render as letters. The woven tag comes out as emboss, or Nano Banana invents a word onto your client's packaging.
that's it. one clay world, four arguments, one product that survives all four.
The EU AI Act's transparency rules went live on August 2. AI-generated image, video and audio must be labelled and watermarked, and a model has to identify itself as one. Anthropic, Google, Meta, Microsoft, Mistral and OpenAI have all committed.
The watermark half is already done for you — Google says every Nano Banana image ships with SynthID embedded. The label is the work: which line of the ad carries it, at what size, on which placement. Decide that before the creative, not after it's approved.
$300k in two months for an AI meal planner. The app writes a weekly menu and a grocery list, and three other meal planners landed in the same store inside the same two months.
So the product is not what separated them. The TikTok output did — the app is what the video sends you to. That's the DTC lesson arriving in software: creative is the channel, not the thing you pick after you've chosen a channel.
🚨New trending app: $300k in 2 months
Mise is an AI meal planner that creates your weekly meals, grocery list and recipes based on your goals
They are going insanely viral on tiktok with these videos:
Ten frontier models math results announced
Five reproduced by a different lab's model inside 24 hours with generic prompt, no internet access, and four of the five by what look like independent arguments.
That is the economics of generated output stated in a single week: whatever one model can produce, another produces a day later without ever seeing the first one's work.
It reads identically in ads. Any panel we render is renderable by anyone holding the same tools by tomorrow.
What isn't reproducible in a day is knowing which of the four panels deserves the budget, and there is no prompt for that, only the result read.
OpenAI unveiled 10 frontier math results from its unreleased Astra model: 24 hours later, an Anthropic researcher says the already-available Claude Fable reproduced five of them.
According to Levent Alpöge, Fable worked autonomously with a generic prompt, no internet access and safeguards against the OpenAI solutions leaking into context.
Only one of the five used essentially the same argument. The other four may be independent proofs.
It seems that OpenAI is now reaping the first-mover success with scientific breakthroughs in new models.
Seedance 2.5 holds 30 seconds in one generation
This is simply insane and takes up to 50 reference assets into it.
The part worth copying from his breakdown is the timestamped storyboard. A single 30-second generation has no cuts inside it, so every beat change has to be written as a time range in the prompt, you are directing, not editing.
The trade runs both ways. The 30-second cut we built yesterday is five 6-second clips hard-cut together: five chances to re-roll one bad beat, instead of one 30-second roll you keep or bin whole.
Long generations buy continuity. They cost you the cut, and the cut is where mistakes get fixed.
i just figured out the best way to use Seedance 2.5
latest model from ByteDance lets you generate up to 30 second continuous videos and accepts up to 50 reference assets (images, videos and audio) all in one generation
most people overcomplicate it, the real trick is in the details
my prompt structure looks like this:
[1] global scene + visual style
[2] character descriptions
[3] timestamped storyboard (0-5s, 5-10s, 10-15s...)
[4] camera movement
[5] negative prompts
[6] ending constraints
for example, instead of saying: "girl doing her skincare routine with x product"
you tell it exactly what happens every few seconds, where the camera moves, how the emotion changes, what should never appear
i put together the exact framework i use and leaking it all to you right now
here's what you're getting:
> the prompt formula for 30-second videos
> how to Use Seedance 2.5 at a fraction of the cost
> the best way to use reference images/videos
> transitions, camera movements and negative prompts
RT + reply "SEEDANCE" and i'll send it over (must follow so i can dm)
4 rules we stick by before a generated panel is allowed to ship:
1. Spell the product out -> cream leather, butter suede, gum sole, one embroidered mark, nothing else. A product described in adjectives is a product Nano Banana gets to redesign.
2. Read the render back with a vision pass and OCR every word on it. The first roll of this sneaker panel came back "the one pair isn't selling you a logo you a logo." Your eye skims that. Gemini doesn't.
3. Wrong text means re-roll the same prompt. Never repair it in Photoshop. Same call, second roll, letter-perfect.
4. Read the argument, not just the spelling. The pass came back asking why the copy says no logo while a mark sits on the shoe, that is your buyer's objection, surfaced before you spend on it.
that's it. The four attached all passed
here's how to stop a generated character standing in front of your headline:
- write the text block into the prompt as a reserved zone, not a position. A box nothing else may enter.
- name where the character stands: a clear side strip, fully outside that box, with a gap between.
- render in Nano Banana, then crop into every text line at 100% and read it letter by letter. At full size your brain fills in the missing one.
- re-roll on the occlusion, never on the vibe — one instruction changed, everything else identical.
The panel below took two rolls. On roll one the figure stood in front of the support line and ate a letter out of two words. Invisible at thumbnail, obvious at 100%.
that's it. the model puts a character wherever it looks natural — usually on top of your second line.
Nobody can photograph being tired. So every supplement ad gives up and opens on the tub.
so I built this one around the day, not the product:
1. the whole film is one depleted day -> alarm, slump, the 3pm nothing. The product doesn't exist for most of it.
2. clay, not photoreal. A sculpted person reads as everyone; a cast face reads as one person the viewer might not be.
3. the box arrives as the resolution of the day, not the subject of the ad. The frame it first appears in is the last thing you see.
You can't shoot a feeling. You can build one, and then let the product walk into it.
Comment "START" and I'll DM you every prompt. (must be following so I can DM)
this can't be real wtf
OpenAI just killed Sora Video Model
We’re saying goodbye to the Sora app. To everyone who created with Sora, shared it, and built community around it: thank you. What you made with Sora mattered, and we know this news is disappointing.
We’ll share more soon, including timelines for the app and API and details on preserving your work. – The Sora Team