Claude chat vs Claude Code, and why Fable 5 low ruined our first build

Can Hoskan wanted to settle something. Everyone talks about Claude Code. Nobody talks about how far the chat window alone gets you, so he started there with the hardest brief he could find.

The brief: a single page tribute site for Messi's World Cup career. 6 tournaments, Germany 2006 through to USA 2026. A Blade Runner 2049 colour palette pulled from 9 Pinterest screenshots. Animated. He ran it in Claude chat on Fable 5 high.

Worth saying who was in the room. Can Hoskan is an Argentina fan, and he picked this brief for a reason. I am a Spain fan. I love Spanish people and Spanish football, and I want Spain to win. We filmed this 2 days before the final, so the bias was running hot on both sides of the desk.

What the Claude chat window can actually build

It came back better than either of us expected. 2 or 3 runs in we had a real page. Sections in a sensible order, the palette roughly where we wanted it, animation doing something reasonable. Promising, and a long way from award winning, which was the question we opened the video with.

That made us curious about the other side. Same prompt, same words, straight into Claude Code.

Fable 5 low vs Fable 5 high on the identical prompt

We picked low effort first because we wanted speed. That choice cost us the entire run. What came back was awful. Not "needs a few fixes" awful. Delete it and start again awful.

We ran the identical prompt on high. Completely different product. Can Hoskan looked at the landing section and said he preferred it to everything before it, including his own chat window version.

Same words. Same reference images. Same tool. The only variable was the effort setting, and the distance between the 2 outputs was orders of magnitude. If you pick low to save 3 minutes, know what you are handing over.

Why the brief mattered more than the tool

Here is the part I set up before we started filming. I wrote that prompt the way stakeholders write briefs. Broad. Confident. No interface direction at all. The only references I gave were photographic. Nothing on layout, spacing, type scale, or hierarchy.

That is what lands in a designer's inbox most weeks, and it produces weak work whether a human or a model reads it. The model filled the gaps with its own defaults, and defaults belong to nobody.

By run 3 the site was decent. Spacing was off in a few places. The direction that got it there was ours.

Use chat for chat, use code for building

The clearest lesson from the day. The chat window is for thinking. Talk through the concept, argue about the palette, pressure test the structure, get the plan right. The moment you want a real product on your machine, move to Claude Code and point it at a folder. Files on your computer, version control, branches you can compare. The chat window has no idea what your folder looks like, so it rebuilds the world every time you ask for a change.

Plan first, execute later. That order removes most of the rework, and it costs 5 minutes.

Keep a progress log so the next session starts where you stopped

This one did not come up in the video, and it is the thing I would fix first.

After a build session, ask Claude Code to write a progress file into the project folder. What was built, what changed, what broke, what comes next. Then next session you point it at the same folder and say carry on from the log.

Ask for HTML rather than markdown. Double tap the file, it opens in the browser, you read it in 5 seconds. No going to GitHub, no digging through commits, no reading raw text. The whole history of the project sits in the folder as a page anyone on the team can open.

The habit: end every session with "write today's changes into the log file". Start every session by asking what the log says.

What designers should take from AI website builds in 2026

After version 3 I had 15 things I wanted to change. That list is the value. Reacting to something real is faster than facing an empty canvas, and all 15 notes came from 24 years of looking at interfaces.

The bar moved this year because everyone has these tools. The output stopped being the thing that separates people. The direction still does, and that part has no setting to turn up.

Plan in chat. Build in a folder. Log everything. Bring your own taste.