Morning Brief
‹ Brief
DTC Podcast · 2d ago · Marketing & Ecommerce

Ep 621: Anything is possible now – The AI Creative Stack

TLDR

🃏 Fable: the model that builds Minecraft from one sentence

Right then. Braden from Pilot House sat down with Eric for a Friday AI session, and the first thing on the table was Fable. What is it? Not a separate app you click into. It's a model inside Claude. Specifically, it's a tamed version of Mythos, the powerful Claude release they were locking down because it was, in Braden's words, basically good enough to hack into systems. Fable is Mythos with the safety guardrails bolted on.

The pitch: it's better than Opus 4.8, and it's stupidly capable. Braden watched a bloke one-shot a fully playable Minecraft game. The entire prompt was "make me Minecraft," then "improve this prompt." An hour later, working game. That's the claim, anyway.

Braden's own use case was sharper. He gave Fable one prompt: rewrite my landing page eight times, spin up sub-agents as a council (an expert copywriter, the CEO, a random customer, a CRO expert), have each rate the eight versions, then a final decision-maker picks the winner. It landed on an advertorial-style page with all the copy written. He took that as a wireframe into Claude design, ran it through the client's existing design system, and got the finished page. His verdict: you could do this with Opus, but Fable feels like another level, especially for web design. His advice is to go search Fable on X and look at what people are shipping.

Here's the catch, and it's the bit you actually need to write down. Claude tells you Fable draws down usage twice as fast as Opus, and it's only included in your plan until June 22. After that, no subscription access. You hand Claude your credit card, click Fable, and it bills you per token, straight, outside your allotted plan. So the free window is closing. People are burning Fable credits on their max plans while they still can.

One honest workflow tip from Braden: Claude eats tokens once a chat's context window gets big, because it re-reads everything to answer the next prompt. His fix is to ask the old chat to summarize itself into a markdown file, then carry that into a fresh chat. Saves you bleeding tokens on one all-day conversation.

🎥 Higgsfield: one founder ad, built between 2pm and 4:30

The platform Braden could not stop raving about was Higgsfield. Think of it as a model aggregator. Instead of paying separately for 11 Labs, VO3 (via Flow), Kling for the video team and the rest, Higgsfield bundles the lot under one roof: Nano Banana Pro, Nano Banana 2, the new ChatGPT image model, SeaDance 2.0 video, the Google video models, 11 Labs, all of it.

The magic, per Braden, is how seamless it is to move between them. Make an image, one click jumps you to video, you prompt your scenes, split into multi-scene if you want, choose 4K or 1080p, pick any aspect ratio (9:16, 1:1, 4:5, 16:9). You can set a starting image, prompt a second image as your ending scene, and stitch a clean output. He took a client's low-fi character from 2018, gave it a Pixar sheen, then built multi-scene video off it.

The headline example: a client joked it'd soon be possible to recreate their founder. Braden hadn't tried, so he just did it. Grabbed a photo, used Higgsfield's "marketing studio" feature, uploaded the product info, images, and a creator avatar (you can spin up infinite AI avatars, target say postmenopausal women, whatever you need), connected it to the product, and prompted a 30-second explainer. Founder in his office, close to camera, B-roll cutaways, slow zoom-ins, him putting on the mask. Then he trained the founder's voice on 11 Labs and laid it over the lot. The call ended at 3pm. By 4:30 he had a believable 30-second ad. The founder's mind was blown.

⚠️ New York's synthetic-human fine and the Andromeda rule

Now the cold water. Eric asked if those ads run in New York. They don't. New York brought in a law, effective the 9th, that says if you run synthetic AI characters or people in ads, you must disclose it. The fine is $1,000, rising to $5,000 for repeat offenders. Pilot House is auditing accounts to find which ones are running AI avatars and working out how to disclose: copy line, or a big glaring label on the ad. Nobody's fully sorted the best practice yet.

The smarter takeaway is don't pretend. Braden's underground DTC circles say you don't need to pass an AI person off as real. Lean into claymation or Pixar style, be openly fake, but tell a real story and offer real value. The line that matters: it doesn't matter if it's human or AI, as long as it isn't a scam. Test both.

Two more nuggets. On AI search (GEO, AEO, whatever Google's calling it), a fresh report says 58% of answer-engine answers are still sourced from cheesy listicles. Funny, because listicles are dead easy to fake, but they still dominate. You still need authentic social mentions, though.

And the big one: Meta's Andromeda algorithm ranks creative on two axes, relevance and targeting, plus how much it affects user experience. Ruin the scroll and it demotes you. Improve the experience and you get a positive vote. That's why socially-native ads work: strip the buttons, kill the obvious red CTA, format for the feed not as an ad on the feed. Pilot House sees it constantly. When the organic team flags a post that's working, they pile spend behind it, and it usually flies.

People have been on Facebook for years. They don't need a button to know they can click.

Watch on YouTube › ‹ Back to the brief