Every prompt from the build, including the vague ones.
Seven prompts took a made-up pottery studio from nothing to a live, deployed, scroll-driven site in one afternoon. They are all here, in order, with a label on each one saying whether it was typed in the build or written for you afterwards.
See the site it built · The sourced write-up
What the labels mean
- Typed in the build — character for character what was typed in the build session.
- Same ask, written out — the same instruction, written properly. The original was a screenshot or half a sentence in chat.
- Not run on camera — never run in the build. Something the video tells you to go do, written so you can paste it.
The build, in the order it happened
Seven prompts took a made-up pottery studio from nothing to a live, deployed, scroll-driven site in an afternoon. This is the whole sequence, including the two that only exist because something went wrong.
The opening instruction
Typed in the build — Cold start. Nothing existed yet — no brand, no copy, no images.
You can make up the branding, you can generate the assets. Two distinctly different sites. Subtle but premium scroll driven animations, text reveals, layering so the whole site has an element of depth and different layers as you scroll, different animations and different depth to make it feel really premium. Relevant to a solo business owner. Maybe scroll animation. Using Higgsfield assets.
Short, and deliberately not a spec. It names a feeling and a constraint and hands over the rest. Notice what it does not do: no page structure, no section list, no colour palette, no copy direction. Those came out of the next prompt, which is the one most people skip.
Make it plan the feelings before it plans the page
Same ask, written out — Immediately after the brief, before a single line of code.
Before you write any code, answer these eight questions in the brand's own voice, and show me your answers: 1. The vibe in three to five words, plus two or three references. 2. The scroll journey, told as the owner would tell it. 3. The energy curve — where it is calm, where it is loud. 4. What the visitor should feel at each stage, and the ONE moment the whole page is built around. 5. One thing this site does that they have not seen another site do. 6. How far from premium-minimal we are going, and in which direction. 7. One unbroken world, or distinct scenes? 8. What assets already exist, and what has to be generated. Then give me the act list before you build anything.
This is the prompt that separates a template from a page. It is answering "what should someone feel at the kiln" before it answers "what goes in the hero." The answer to question five became the site's signature move: scrolling throws a pot in the corner of the screen, and opening the kiln glazes it.
The image style preamble
Typed in the build — Pasted at the front of every single image and video generation.
Cinematic documentary photography of a solo potter's studio, shot on 35mm anamorphic lenses. Shallow depth of field, true blacks, matte film grain, high dynamic range. Low-key lighting: one cool north-window daylight key, deep falloff into shadow, a single warm practical lamp. Colour grade of deep slate-green shadows, bone-white highlights, desaturated mid-tones, iron-red and ash-grey clay. Photographic realism. NOT 3D render, NOT claymation, NOT illustration, NOT CGI, no digital glow, no plastic sheen, no text, no logos, no watermark.
This is the reason twelve separate generations read as one photo shoot instead of twelve stock images. It is reused character for character every time; only the subject line after it changes. The negatives at the end are doing as much work as the description — "NOT claymation" exists because a pottery prompt goes there on its own.
A single shot request
Same ask, written out — Twelve times, once per still, with the preamble above in front of it.
[paste the style preamble above, unchanged] Subject: a kiln door standing open in a dark studio, three shelves of pots glowing orange from the heat inside, the room around it almost black. Camera at chest height, straight on, the glow the only light source. Nothing in frame that reads as modern.
One subject, one camera note, one exclusion. That is the whole shape. Of the fifteen generations, twelve made the cut: one clip was unusable, one still was off-brief, and one was rerolled. All of it cost about eighteen credits — a couple of dollars — which is the number the video quotes.
The refinements
Typed in the build — Three separate rounds, each after looking at the actual page.
smoother scrolling, text needs more time to breathe the text is too abrupt make it as beautiful as you can
Three of them, hours apart, each one vaguer than any prompting guide would allow. They worked because the model had its own written brief to measure them against — it knew the argument section was supposed to feel like conviction, so "too abrupt" had somewhere to land. The last one produced the giant pot turning behind the argument text and the ember wash across the kiln. This is what long-horizon actually looks like: not one magic prompt, an afternoon.
Make it check its own work
Same ask, written out — After every round of changes. Five times in total.
Screenshot the page at 49 scroll positions, at desktop width and at phone width, and lay them out as a contact sheet. For every frame, check that text over imagery clears 4.5:1 contrast, that nothing is caught mid-animation, and that no screen is dead scroll. Then read the sheet back to me and tell me what is wrong — including anything the automated check passed but looks wrong to you.
The most valuable prompt on this page, and the one nobody puts in a demo. The last sentence is the important half: the automated check went green while the vase was rendering invisible behind another layer, and only reading the sheet by eye caught it. Two of the bugs in the video were found here, by the model, in its own work.
Ship it
Typed in the build — Once, at the end.
npx wrangler pages deploy dist --project-name halden-clay --branch main --commit-dirty=true
A command rather than a prompt, and the only one on this page. Seventeen files, 5.2 MB, live on a free Cloudflare Pages project. The site in the video is what came out the other end of this line.
The decision the video actually asks you to make
Anthropic's own docs say start with Opus 5 and only reach for Fable when Opus at high effort falls short. This is that test, written so you can run it on your own weekly task this week.
The Opus-first test
Not run on camera — Before you spend a dollar on the more expensive model.
I want to test whether this task needs a more expensive model than Opus 5. Here is a task I do every week: [describe it in two or three sentences — the newsletter, the client report, the podcast-to-clips grind]. Here is the raw material: [paste or attach it]. Run it at high effort. When you are done, tell me plainly: - which steps you were confident on and which you were guessing at - anything you had to invent because the input did not cover it - what you would need from me to do this without supervision next week
Run this on Opus 5 first. If it nails the task, you have your answer and you just saved yourself Fable money. If it misses steps, invents things, or needs babysitting, run the identical prompt on Fable 5.1 and compare the same three answers — now you know exactly what the extra money buys on your work, rather than on a benchmark.
The other two jobs from the video
Neither of these was filmed. Both are things the video tells you to go do, so they belong here rather than in a paragraph you have to reconstruct from memory.
Read my own site for security problems
Not run on camera — When you are running code you did not write and cannot read.
Review this site's code for security problems. I'm not a developer. For each issue, tell me the file, what's wrong in plain English, what could happen if it's left alone, and how to fix it. Do not write any exploit code — defensive review only.
Fable 5.1 will do this now; earlier versions refused outright. Two limits worth knowing before you run it. It still will not write attacks, and asking for them reroutes you to a different model. And point it at code you own — your own site, your own theme, your own plugin. Not somebody else's live site.
The email to your AI vendors
Not run on camera — This week, if you pay for any tool built on Claude.
Hi — quick question about the Claude Fable 5.1 release on September 1. Cache reads dropped 75%, and Cognition reported the same Devin task falling from about $5.00 to $2.68. If your product runs on Claude, that is a real change to your cost per job. Is that showing up on my side as a lower price or higher usage limits, or are you keeping it as margin? Either answer is fine — I would just like to know which one it is.
Not a prompt, and the only thing on this page that involves another human. It is in the video because it is the one action that turns somebody else's price cut into your money. A vendor who answers this straight is worth keeping.
Where this sheet stops
- These prompts are shaped by a very specific job — one scroll-driven marketing page for a fictional brand, built by someone who has been paid to build sites before. Paste them at a different problem and you will get a different quality of answer.
- Prompts two, four and six were reconstructed from a build log rather than copied from a chat window, because the originals were a screenshot and two half-sentences. The instruction is the same; the wording is tidier than what was actually typed.
- The build ran in Claude Code on Fable 5.1 with a project skill loaded for scroll-driven pages. The eight planning questions come from that skill. Without it, prompt two is still the right move, but you are asking the model to invent the questions as well as answer them.
- One afternoon on one project is not proof of anything. It is one honest data point, and it is the only one this page can offer.