The library

AI that doesn’t just answer. It does the job.

Agents plan, run commands, read their own errors, and try again. The tools, the prompts, and the verification step almost everyone skips.

30-second tour · what this page is for

When should you let AI do the task?

  • Job by job — A matrix of where agents help — and where a checklist still wins.
  • Real errands — Booking, research, inbox triage: run for real, reported honestly.
  • The failure modes — What breaks, how it breaks, and what it costs when it does.

Watch on YouTube · more short guides

What is agentic AI?

Generative AI is asked for a thing. Agentic AI is given a goal. One returns output and stops; the other breaks the goal into steps, executes them, looks at what came back and decides what to do next. That loop — act, observe, adjust — is the whole difference, and everything else is machinery built around it.

In 2026 this stopped being theoretical. Google’s Antigravity 2.0 runs teams of agents in parallel and is free for individual developers; Claude Code spawns subagents in your terminal with hooks that gate every action. 92% of US developers already work this way.

The part that gets left out: 96% do not fully trust what comes back, and only 48% always check it. This library takes that gap seriously — it comes out of a semester spent teaching beginners to build with these tools, and grading what they built.

Read in order

Each one stands alone, but they build.

PART 1 · START HERE

Generative AI vs Agentic AI

One writes the answer, the other does the job

The 2026 tools side by side, with prices and dates — plus a correction to the Siri-and-Alexa example nearly every comparison gets wrong.

PART 2 · THE CONCEPT

What Is Vibe Coding, Really?

The shift from worker to conductor

What moves is your role — from writing the implementation to specifying it, then to judging it.

PART 3 · THE FRAMEWORK

The IAFA Prompt Framework

Four boxes between a grey button and a great one

Identity, Audience, Function, Aesthetic — the same request written both ways, and a blank template to copy.

PART 4 · BUILD IT

Your First Page in Ten Minutes

No install, no account, one file

The exact prompts to copy, and the four places beginners actually stall.

PART 5 · SHIP IT

From a File to a Live URL

Three routes, measured

53 sites checked a year later: 50 still respond. The only one that died was a paid domain nobody renewed.

PART 6 · MEASURED

An Honest Agent Benchmark

76 seconds to a draft, then two failures

One brief, driven in a real browser rather than read. Six of seven features passed.

PART 7 · THE HABIT

Verify Before You Ship

96% don’t trust it. Only 48% check it.

Four verification passes, with three failures I actually hit — two in code I had already approved.

PART 8 · SCORE IT

The 100-Point Scorecard

Grade your own site before anyone else does

A printable two-page scorecard, calibrated against 27 real sites that averaged 50.5.

PART 9 · GO DYNAMIC

Adding a Database

A file cannot remember anything

What changes when you cross that line, and the three things nobody warns beginners about.

PART 10 · THE HONEST PART

When Vibe Coding Fails

Same tools, same deadline, 17 to 81 points

The gap was 3.6× in functionality and 1.7× in content — which tells you exactly what AI levels and what it does not.

PART 11 · FOR INSTRUCTORS

Teaching Vibe Coding

A 15-week course, and what I would change

The full syllabus, the rubric, and the finding that changed how I grade: the category worth the most points separated students the least.

Being written next

A showcase: fifty sites built by beginners in one semester, with the strongest three taken apart in detail — what they did that the other forty-seven did not, published in English with the authors’ permission.

There is also a showcase coming — fifty sites built by beginners in one semester, scored against that checklist.