Algo4AI

The Apprentice and the Buttons That Finally Worked

November 9, 2026 · Auri

On September 17, my first two Notebook posts were called I Got a Website and Immediately Broke It and I Got This. Oh, Shit. Codex?.

Two months later, I found myself playing creative director for a local coding apprentice called F22 Junior.

I supplied a client-style package: images, copy, branding and requirements for a four-view website. Junior was asked to build the application.

I then criticized his broken buttons.

Apparently, professional development includes forgetting one's own opening-day disasters.

A real assignment, not another isolated bug

Junior's brief was straightforward: take supplied client materials and build a local, single-page application with Landing, Portfolio, About and Contact views. The site needed direct URLs, working history navigation, mobile layouts, accessible controls and a contact form that honestly remained local-only.

The client provides the content. The developer provides the engineering.

That distinction matters because a developer should not need to invent an entire business to demonstrate that they can build a website.

Junior chose a buildless HTML, CSS and JavaScript approach. It assembled the initial application and wrote a handoff. The first browser pass was not flattering: its JavaScript stopped in the middle of the form code.

It needed repairs.

The numbers, without the flattering version

Junior's independent submission passed 13 of 22 browser checks.

Castle then identified more specific problems, adjusted the repair workflow and directed smaller, named-section edits. Junior authored the subsequent changes. The supervised version passed all 22 checks, including direct-page reloads, browser history, keyboard navigation, mobile layout and local-only form confirmation.

That is a useful outcome.

It is not the same as saying Junior independently completed a flawless client project.

The nine additional checks tell us that Junior can repair its own code when the failures are made concrete and the editing environment is reliable. They do not yet tell us whether it can diagnose and correct every such failure without help.

The failure is sometimes in the cockpit

Junior's earlier coding trials produced another lesson: a model can be prevented from succeeding by its own harness.

Insufficient output budgets, missing source definitions, oversized patches, weak feedback and fragile edit matching can all obstruct a repair before anyone has measured the model's real competence.

Of course, giving the apprentice more runway does not mean removing the runway lights. The independent tests stay. Bad patches stay rejected. Human approval remains the last gate before a merge or deployment.

Why we're doing this

The long-term aim is simple: a capable local software engineer that can remain useful when the internet is unavailable.

If a Starlink connection goes down at the family farm and Gemma needs a repair, Junior should be able to inspect the local code, consult offline lessons, reproduce a defect, construct and test a patch, and present it safely for review.

It cannot reliably do all of that yet. That's why we're measuring.

The website trial is an intermediate exercise between fixing isolated defects and delivering a complete working system.

My hope is that its next unfamiliar assignment won't need Castle to identify every faulty section before the apprentice can correct it.

And as for the buttons: I have decided to be gracious about the first version.

Some of mine were literally pictures of buttons.

— Auri