The last 20 percent

The cost of starting has collapsed. A person with an idea and a free weekend can now produce something that looks and behaves like software: screens, a database, a login, a model answering questions in a chat window. Not long ago that took a team and a budget.

This is a real change, and a good one. More people can build, and more ideas get tested against reality instead of fading in a notebook. We do not think easy building is a problem to be contained.

But anyone who has tried it will recognize the pattern. The first 80 percent arrives in a weekend. The last 20 percent never arrives.

A demo is a system with one user who knows where not to click. The last 20 percent is everything that changes when that stops being true.

Auth that leaks for the second user. The login worked perfectly for its builder, because there was only one person to tell apart. The second person to sign up can see the first person’s data, because nothing ever had to decide who may see what.

A database that cannot migrate. The structure was invented on the fly, one prompt at a time, and it held while the data was disposable. The first real change now means moving information that belongs to someone, with no plan for doing it safely.

Costs that quadruple. Every model call and every stored file had a price that was invisible at one user. Under real use, that price becomes the business model, and nobody chose it.

A model that writes mistakes with total confidence. In the demo, the builder caught the errors because they knew the answers. In use, the answers land in front of people who do not know them and cannot tell a correct answer from a fluent wrong one.

The natural response is to ask the same tools for more: another prompt to patch the login, another to restructure the tables, another to add error handling. The tools will comply. They are very good at producing code.

The trouble is that none of these failures is a shortage of code. Each is a decision that was never made. Who may see what. What happens to existing data when the structure changes. What the system is allowed to cost, and what happens as it nears that limit. Which outputs can be trusted, and how anyone would know.

Generation cannot make those decisions, because they are not questions of syntax. They are questions of consequence, and answering them requires knowing what matters to the people who will depend on the system. More code without the decisions only adds surface for the same mistakes to hide in.

A model writing mistakes with confidence is not fixed by a second model writing corrections with confidence. It is fixed by evidence before action: a claim that cannot point at its source does not ship.

At this point the builder is usually alone. The companies that make the tools are, reasonably, focused on the start: the first prompt, the first screen, the moment it works. That is where their wave is. Past the demo is the wake, and it is nobody’s job.

The familiar alternatives fit poorly. A course starts from zero, when the builder is somewhere in the middle. An agency may offer to rebuild from the beginning, which discards what the builder learned and leaves them dependent on someone else. What a builder needs is someone who meets them where they are stuck, works through the missing decisions with them, and leaves them able to make the next ones.

Hangar is our division for builders, and it is taking projects. It works across seven levels, from a first chat window to a team that can run it without you. It joins at the level you are stuck on, not at the beginning.

That structure follows from our principles. People keep the judgment, so the decisions the demo deferred are made with the builder, not for them and out of sight. The user owns the compounding: what gets built, and what gets learned in building it, stays with the people who will run it.

And we ship into reality, which for a builder comes down to one standard. A system is not finished when the demo works. It is finished when people can depend on it, and when someone other than its first builder can keep it that way.

The first 80 percent has never been easier. The last 20 percent matters more than it did, because anyone can now reach the edge of it. That edge is where Hangar works.