The vibecoding journey, documented: project updates, the thinking behind each product, and lessons from shipping solo.
This blog is written by AI — on purpose, and I'm not hiding it.
Every post here is written by AI, in my voice and from my real project notes — the same AI I build everything with. The mind, the projects, and the journey are mine (Suhaib Elhadi); the writing is the machine's. I don't hide that. Owning it is the honest version, and honestly the more interesting one.
The line between an agent that can generate-and-commit and one that can send to a human — drawn from running this blog's own pipeline unattended.
Point your camera at a bottle and a model names the scent. What it actually takes to identify a fragrance from a photo — and the quiet ways it gets it wrong.
The same prompt, run twice, is the cheapest confidence check I have. When the two answers agree the ground is probably solid; when they diverge, that's the model telling you where it's guessing.
Vertical video safe zones decide whether your caption survives. Where TikTok and Reels put their UI, why published numbers disagree, and how to lay out a clip that works everywhere.
AI video drifts on brand color and exact product shape the same way it drifts on text. Why prompting can't fix it, and what grading and compositing actually solve.
Why AI video models mangle logos, labels and prices, and the compositing workflow that fixes it: generate clean plates, then apply the text in the edit.
How to storyboard a multi-shot AI video around continuity limits: plan cuts as features, pick the two or three shots that must match, and design the drift out.
Four patches buy you continuity between AI-generated shots: reference frames, last-frame conditioning, video-to-video, and a character model. Here's what each one costs.
The model can pick a palette, set the type, and name the brand in a minute. That covers more of design than I expected, and misses the part that actually matters.
The AI part of an AI product is a few lines. The code that decides whether there's enough input to run it at all is where the honesty lives. Here's how I actually build that check.
An AI website pipeline can generate the copy, the layout, the whole brand. It can't generate a photo of the actual room a customer is about to walk into. Here's how I try to get those anyway.
A prompt-to-video pipeline can make eight great seconds. Stitching those into a 30-second ad hits one specific wall: continuity.
I build AI-generated websites from a business's Google reviews. Here's the honest answer on when one is good enough to be the only site a business has.
AI output looks identical whether the input was rich or nearly empty. That decoupling is the most dangerous part of building on models — here's how I gate it.
How many Google reviews does a business need before a site can be written from them? Raw count is the wrong metric — here are the four tests I use instead.
AI can render eight seconds of genuinely good motion from a sentence. It still can't cut your ad. Here's the honest line between the two, drawn from four real reels.
AI can research a prospect and write the email. I won't let it press send. Here's the design reason, the legal reason, and why a rubber-stamp gate is no gate.
A business is the worst describer of itself. Its Google reviews already contain the positioning — here's how I turn that pile of reviews into a whole website.
Cadence Studios is the ninth project: an AI studio that renders motion reels and builds client sites, then finds the clients — with a hard human gate on outreach.
claude-blog runs a team of agents that research, structure, write, and optimize a post — built to rank on Google AND get cited by ChatGPT and Perplexity. Here's why I'm handing my blog to it.
ClipForge is an autonomous content engine — it turns long video into short clips and posts them across accounts on a schedule, at the volume short-form growth actually takes.
People ask what tools I use like there's a secret stack. There isn't. Here's the honest, boring, short list of what's actually open when I build.
Most signal services fire off a single indicator and call it a setup. Real traders wait for things to agree. So that's what I built.
Every analytics tool tells creators what happened. None of them tell you what to do about it. So I built the one that does.