The word that has stuck in the trade for how we do this is vibe-coding. It landed as a joke and is now a term of art, and half its meaning has inverted. Originally it meant point the model at an idea and hope. What is actually happening on the machine in front of me is different, and it has a better name. I call it guide-coding.
In guide-coding the model writes most of the lines. I own the decisions it should never make for me. I set the architecture, I name the boundaries, I review the diff, I run the build, I take it to review. The model is a superb hands-on. The steering stays human. That single division of labor is the whole thing, and it is why the output is software a stranger can install instead of a demo that dies the moment you touch it.
Two Apps, One Method
The cleanest proof I have right now is the two native apps sitting on my Mac. DrawmaticAI turns a plain-English description into a production diagram. Architectures, flowcharts, ERDs, sequence and network diagrams, generated by a local model over a bundled, offline copy of the open-source draw.io editor. OwnScribe, the third and newest, takes live speech and writes clean minutes: Whisper transcribes on the GPU while I talk, then a local LLM cleans the transcript, summarizes it, and pulls the action items. Both run entirely on Apple Silicon. Both are guide-coded in the same afternoon-to-review window.
Watch what that means in practice. I did not draw a diagram in DrawmaticAI and call it a model. I specified what a good diagram looks like, wired the model to the canvas, and shipped it. I did not wire up OwnScribe's microphone, the waveform, the encrypted store, the Model download, the App Store category, the entitlements, the sandbox. I decided which of those must be local, which must be encrypted, and what the machine is allowed to touch. The model produced the code for all of it. I produced the parts of it where being wrong is expensive.
Who Actually Benefits
Everyone who is not a staff engineer with an infinite budget, and that is most people who need custom software. This is the part the hype misses.
One person can now hold a portfolio. The MyOwnLifeOS suite is not a team's output. It is mine, one reviewer, with the model doing the mechanical writing. That used to be impossible. It is not that the code is free. It is that the marginal cost of the next app dropped below what a single human can review in a day, and the human is exactly the ceiling we want to keep. When the machine can write a native macOS app from a spec, the bottleneck flips from typing to judgment, and judgment does not scale, which is the good kind of ceiling.
Small businesses get to own their tools. A practice with fifty people has always had a choice: rent a cloud suite someone else built and pays per seat, or commission custom software that costs six figures and takes a year, and then still send their most sensitive data through that vendor's API. Guide-coding opens a middle door. You can describe the tool you actually need and, because the stack runs locally, keep the data in the building. MyOwnLifeOS apps are built that way. That is not a feature I added. It is the only way I could build it and keep it mine.
Privacy stops being a premium tier. The reason local inference felt like a hobby was the cost, and the cost has just collapsed. Apple Silicon runs a real language model and a real speech model on the integrated GPU. So a person who could not afford a GPU tower can now own the private option by default. The people who benefit most are exactly the people who can least afford to leak: therapists, researchers, operators handling other people's data. They get sovereignty without the capital stack.
The Human Is Still the Point
There is a version of this story where the model is the smart one and the human is a button. I have watched it fail, and I would not ship it. The failure mode is not broken code. It is correct, confident software doing a thing nobody decided should be a thing. The scope bloat, the silent data collection, the over-engineered thing that is also slow and also a risk, all of that comes from a missing human at the decomposition step.
Guide-coding is the discipline of not letting that happen. The model is allowed to be fast and wrong. I am the one who is not allowed to be wrong, so I stay on the parts where wrong is costly, and I let it do the rest. The apps ship lean on purpose. Fewer features than the trend would suggest, no dark patterns, data that does not leave the device. That restraint is the human showing up.
The future is not a model that writes the whole app. It is a person who can direct one, who can judge its output, and who chooses to build local because it is now affordable to. Guide-coding is that person with a very good pair of hands.
So when I say guide-coding is the future, I mean a future where owning your software is not a company-size decision. Where the private option is the default because it is the cheap one. And where the person in the loop is the product, not the bottleneck. DrawmaticAI and OwnScribe are the receipt. Two native apps, one owner, local to the last byte.