The loop in depth

Module 3 is the four moves you make with your agent, one lesson each: know what you want, ask, look at what came back, say what should be different.

Every lesson runs the same worked example on one practice page — a page your agent creates for you in Lesson 1 and you watch change in your browser after every step. Exchanges are shown from a real Claude Code desktop run. The ChatGPT app with Codex takes the same sentences, but its replies were not recorded for this module — you run only the app you picked in Module 0, and check the same page either way.

What this module builds

By the end of this module, you can run the four moves with any agent: knowing what you want, asking for it well, judging whether what came back matches, and course-correcting when it doesn't.

Each lesson builds on the last:

• Lesson 1 — Introducing the loop: name the four steps in order, have your agent build the practice page from one sentence, and run all four once on it → sets up Lesson 2 by leaving the ask as the step to sharpen first.
• Lesson 2 — Plan before you build: ask for the plan before the work, and know when and how to start a fresh conversation when a long session starts sliding → sets up Lesson 3 by giving you a plan to compare your agent's actual output against.
• Lesson 3 — Check the plan and the result: run five checks that flag wrong output — including an answer your agent confidently made up — without ever reading code → sets up Lesson 4 by leaving you holding a wrong result you now need to fix.
• Lesson 4 — Steering and recovery: write a three-part steer, catch your agent doing far more than you asked and pull it back, go back to a saved version once on purpose and check what it gave back, and know when a fresh conversation beats another steer → sets up Module 4, where you put the loop to work building the thread project.

The thread that ties it together: the four moves are durable and the app around them is not. The window changes, the buttons change, the agent's name changes. Knowing what you want, asking for it, checking what came back, and saying what should be different — those four carry to whatever you end up working with next.

Worked example

All four lessons move the same practice page forward, one step at a time. There is nothing to download and nothing to install: in Lesson 1 you ask your agent to make the page — a folder called loop-practice with one page in it called index.html, showing your name and a one-line tagline — and to open it in your browser. Saving is the first thing that may need something set up on a fresh computer; Lesson 1's exercise has your agent check and install it.

From there the state chain runs across the module. Lesson 1 adds today's date below the tagline. Lesson 2 adds a button that shows and hides the date. Lesson 3 asks for a list of three favorite books and gets back three the agent made up. Lesson 4 steers that list back to placeholder text and saves it, lets a loose ask put a wooden background on it, goes back to that save on purpose and checks the placeholder came back, then gives the list its look from a fresh conversation and saves the page finished.

The page is throwaway on purpose — a wrong answer on it costs you nothing. Your real project starts in Module 4; by then you can ask your agent to delete the practice folder or keep it as a souvenir.