This week I have built FIVE native macOS apps with Opus 5.5: A notion-like Markdown editor A supercut-like screen recorder A descript-like Video editor A Canva-like image editor A Keynote-meets-Miro whiteboarding/presentation app All one-shotted. What is happening.

I just made my first video with AI - a 15 sec commercial for a made-up drink bottle. Here's everything I learned making it: - Generations are expensive/take a long time and this constraint drives basically the whole workflow (its the opposite mentality to code which is cheap to generate) - That wasn't a big factor on this project as I was using @higgsfield_ai which had a nice promo which gave 33 days of unlimited usage of Seedance 2.5 (the new 'best' model), so I was probably a little more liberal with my video generation than I would have been otherwise - Using images to create reference sheets for characters, locations, objects is a must for keeping consistency between scenes. In Higgsfield you create these as specific labelled reference images that you can @ mention in your prompt - Higgsfield had a nice skill that I copied from their Youtube video for beefing up prompts with technical details like camera lens. I didn't compare with more raw prompts but it seemed to work well - With Seedance 2.5 I was only able generate at 720p but that was fine since I ended up upscaling everything to 4k using Topaz Artemis upscaler later - Seedance 2.5 is generally good at generating entire sequences with multiple cuts but ultimately I found using more specific prompts for each individual shot gave better outputs. The multi-shot clips tended to have more weird artifacts and illogical behavior (like a person walking through a kitchen counter - The general workflow was to use Claude to refine a prompt, paste it into Higgsfield, generate, see what went wrong, refine the prompt, retry, and continue until I had an output I was happy with - Music was initially tricky and I tried a few different tools. In the end the @ElevenLabs EMusic API gave me what I wanted; a JSON scheme describing the track I wanted, broken up in 3sec segments (so I could define the intro, mid section, outro independently), and I could orchestra directly from Claude Code - For my next video, I want to try generating image stills from imagined timestamps in the final shot, for example one of a person entering a restaurant 1 sec in, another of them sitting down at 4sec. Basically blocking out the action in images, a mini storyboard just for that shot. I'm hoping that will keep the outputs closer to what I want on the first and second time (so I don't waste credits) - Higgsfield also have an MCP so I can see many people (myself included) building a little local app to orchestrate everything with Claude rather than copy/pasting prompts into the Higgsfield (though Higgsfield UI is actually pretty great for organising your shots)

Hard disagree on this. Sure, I'll take 100% AI (and I do everyday) for learning/research/updates. But the purpose of 'content' is to feel connected with another human. It can be 99% AI / 1% human. But I need to know that a human has bared at least a sliver of their soul.

I'm beginning to feel that a huge reckoning is coming on multiple fronts once the mainstream understands that LLMs != AI. There is no intelligence in a probability machine for text. All these servers and energy costs and noise for nothing more than a powerful autocomplete.

I was all claude code a few months ago. Last month or two I've been all Cursor. Switched because I preferred the DX and the costs were similar. Now everyone is bullish on claude code again. What I miss?

As a remote worker I'm constantly checking times in different timezones and typing meeting options into emails/slack (sometimes Calendly is overkill). Finally carved out some time to build a tool to streamline this which I've had in my head for years. Thanks Claude.

The debate between vibe coding and pure engineering is missing the sweet middle where you're not some idiot prompting yourself around in circles nor are you obsessing over syntax. It's a middle more akin to being the director of your own movie. You don't need to know what ISO to set the camera to, what bitrate to record the audio for, or which colour gel to put over the lights, but you do need to be an expert in cinematography, storytelling, understand the difference between a dolly and a crane, and be able to coordinate all these disparate parts, these mechanical pieces, that dance behind the scenes to the tune of your unique vision. The AI coding equivalent is expertise in patterns, architecture, managing context, in picking the right tool for the job, and most of all in communication. AI doesn't mean there's no longer anything to learn in order to build software but the things you learn are those that make you sufficiently technical to speak the language of software engineering.

Coming from Bubble/nocode, one of the sore spots working in code has always been finessing the UI. No more! Absolutely loving the way no code and ai-coding are converging.

I'm building a dead-simple security tool for Lovable apps. I'm looking for some beta-testers. Comment 𝘁𝗶𝗰𝗸𝗲𝗱 below for early access.

I create a board of advisors in the Claude mobile app based on real life, successful people I admire and chatted to them in voice mode. They sometimes disagreed with each other and other times were unanimous in their outrage at my suggestions. It was my own private council of mentors. For every question I got different perspectives and could make up my own mind. I've bounced ideas off AI many times before but this was the most realistic and insightful session I've probably ever had. It was using Opus 4 which probably played a role as well.

Easily one of the greatest and most well executed products to emerge in the @bubble ecosystem, EVER. George is absolutely cooking 🔥

Big day! The ability to build mobile apps with @bubble is now open to everyone. Would love to know what ya'll think of the public beta. If you're new to Bubble, I've made a Getting Started course for Bubble's Youtube channel: youtube.com/playlist?list=PL…