See, critique, iterate, ship
Iterate by seeing — the professional loop
Exercise 1.1 gave you the WOW. Exercise 1.2 showed you that contextContextEverything Claude can see right now — the conversation so far, the files it has read, the tool results, the system instructions. A big text buffer the model reads end-to-end before every reply. beats the prompt. Now we go one level further: let the agent see what it built, critique its own output, and iterate.
This is the move that separates amateur from professional use.
Pick ONE artefact to iterate on
- A) Your game — create a game that simulates something in your business.
- B) A dashboard for your domain — real or mocked dataset, charts, filters, interactivity. Something that could plausibly live inside your company on Monday.
Extra
- Design Idea: Learn design language and see your product improve.
- Ecommerce Dataset (original): Complete dataset for building dashboards.
- Prosus Report: Has information about each company (please extract this to .md first).
- huggingface.co/datasets: Biggest collection of datasets (don't download 100 GB!).
Start
Prompt:
Read <name>.md, I want to make .......... use .....
Ask me some questions so we can build this together.
The goal is to make something that shows the power of AI
and I can share with my peers via Vercel.
Their reaction must be "Wauw, did you build this?"
On Repeat
Prompt:
Launch it and examine it. Then reflect on your own
creation and propose three concrete improvements.
OR
Prompt (example):
Check out the button, it's too small to click
The agent sees the artefact (screenshots, DOM, console), critiques it, and proposes changes. You agree with some, reject others. Re-run. See. Iterate.
Do this at least twice. Notice how much less you are typing and how much more is changing.
Deploy it
When you're happy, ship it:
Prompt:
Deploy this to Vercel and give me the live URL
If you did not create a game in Exercise 1.1 / 1.2, use this snake starter: snake example
- Download the html file
- Move it to your Claude directory
- Open it by double-clicking to examine
- Start the loopLoopWhat makes Claude an agent and not a chatbot. Instead of one ask-and-answer turn, a loop runs Claude over and over: act, observe what changed, decide the next step, repeat. above against it
Why we do this: iteration is how real work gets done. One-shots are demos. The professional loop is see → critique → update context → re-run. You are not reviewing the code — you are reviewing whether the agent did what you meant.
Stuck?Debrief — what should have happened
A deployed URL that looks substantially better than version 1. Crucially — you barely typed. The work happened in the loop between agent and artefact. You stepped out of the metronome.
Forget the code even exists. — Head of Claude Code
The point of this chapter: iteration is how real work gets done. One-shots are demos. The professional loop is see → critique → update context → re-run.
Stuck?FAQ for this exercise
Q: Can you share how that level of design quality was achieved in the demo? A: Ask Claude Code to ask you questions, take references from other websites, act as an Apple designer, make it as simple as possible. Then add a few touches like a dark mode toggle.
Q: Can we send Claude a folder of files and say "go through each file and create an md file for us"? A: Yes. Open a Claude sessionSessionA single ongoing conversation with Claude Code. Every message, every file Claude has read, every tool result lives inside one session — stored locally on your machine. in that folder and ask: "explore all the files in this folder and generate an executive summary." It will read them and produce the .md.
Q: What's the difference between Co-work and Claude Code? Today I give Co-work access to a folder and get a lovely summary — should I use one for one case and the other for another? A: Co-work has restrictions; Claude Code can do everything. Co-work edits exist only inside the project conversation session, so it's confusing whether edits are "real" or "fake". Claude Code changes your actual files in the directory. For this course we use Claude Code because we need that power. Some workflows can eventually move to Co-work because it's easier and safer.
OptionalGo deeper
Mini-hackathon — pick ONE track. The loop is the same; the artefact changes. Run see → critique → iterate end-to-end on whichever track grabs you:
- Track A — Slide. Generate a single Prosus-styled slide on a topic you care about. Screenshot. Ask Claude to critique against brand and content. Iterate until it'd survive an exec review.
- Track B — Email. Draft a tone-checked email using your
voice.md. Render to PDF or screenshot the preview. Ask Claude to critique against the voice file. Iterate until it sounds like you. - Track C — Dashboard. Build a one-page metrics dashboard (mocked data is fine). Screenshot. Ask Claude to critique against a Grafana board you actually use. Iterate until it'd plausibly live in your team's wiki.
Whichever you pick, the lesson is the loop, not the artefact: see → critique → update context → re-run. The format is incidental.