AI that does not just answer — it plans, executes and checks its own work. Tools, workflows and the verification nobody talks about.
Agents win where a machine can check the machine. Assistants win where the only judge is a person. The matrix, the half-life maths behind it, and three sets of numbers – including our own timed runs.
The full fifteen-week syllabus, the 100-point rubric, what 27 graded submissions actually produced — and the four things I got wrong, including a rubric whose point weights turned out to be backwards.
27 people, identical tools, one deadline — and a 64-point spread. The gap was 3.6× in functionality and only 1.7× in content, which tells you exactly what AI levels and what it does not. Five failure patterns, what the top group did differently, and why none of it was talent.
Of 27 sites, 15 connected a database and 12 stayed static — and the widest gap between the best and worst work was exactly here (3.6×). What changes when you cross that line, the prompt to ask for it properly, and three things nobody warns beginners about — starting with the key everyone can see.
“It looks fine” is not a score. The four-part rubric I grade with — design 30, functionality 20, content 30, accessibility 20 — plus what 27 real sites averaged against it (50.5) and a free two-page printable scorecard.
96% of developers do not fully trust AI-generated code. Only 48% always review it. Four verification passes — drive it, break it, check what you never asked for, and suspect the test itself — with three failures I actually hit, two of them in code I had already read and approved.
I checked 53 sites deployed on free hosting a year later — 50 still respond. Three routes to a live URL compared by what they assume you know, the five things that will bite you (including error messages that are not errors), and when a custom domain is worth it.
A text editor and a browser is all you need. Write four lines of brief, ask for one complete file, save it, then change it by typing sentences. Includes the exact prompts to copy and the four places beginners actually stall — none of which are the model’s fault.
76 seconds to a working draft, then two failures. One was the agent’s — a scroll trigger five percent tall that reading the code would never reveal. The other was mine: a test environment that could not produce the event the code was listening for.
Ask for “a cool cyberpunk button” and you get a grey rectangle. Ask in four labelled boxes — Identity, Audience, Function, Aesthetic — and you get the thing you pictured. The prompt structure I teach beginners, with the before-and-after, a blank template to copy, and the three mistakes I see most.