Most Claude Code examples show a polished end result and skip the prompt. These do the opposite: here is exactly what we sent, word for word, and exactly what came back, once, unedited, from Claude Opus 5.5 and Claude Fable 5.1. Start with the shortest prompt of the eight, which produced the most surprising results; its opening lines are quoted below the builds.
Invent a product and build its website.
It can be anything: a physical product, an app, a service, a place, a food, a tool. You choose the name, the idea, the story, the visuals and every word. The only requirement: make it memorable.
That is the opening of the prompt, word for word. The rest sets the page requirements: a hero, at least three more sections and a footer, real copy throughout, a layout that works at 1440 by 900 and at 390 by 844, and every visual made with HTML, CSS, SVG or canvas. Read the full prompt.
What did each prompt ask for, and what came back?
Eight prompts, the gist of each and the two results. The full text of every prompt is on its brief page with a copy button.
- Landing page. Asked for: a full marketing site for a scheduling app called Driftwood, with fixed copy facts and prices and a monthly-to-yearly switch. Came back: a hero with a working booking calendar (Opus) and a single serif headline over a booking animation (Fable).
- Analytics dashboard. Asked for: an admin dashboard on a supplied dataset, with KPIs, a hand-drawn chart, a sortable and filterable table and dark mode. Came back: correct figures in both, with an annotated peak (Opus) and a chart-to-table switch (Fable).
- Physics simulation. Asked for: twenty balls in a hexagon turning once every ten seconds, physics from scratch. Came back: two stable simulations; Fable's is the smallest file on the site at 15 KB.
- 3D shooter. Asked for: a three.js arena with waves, recoil, HUD and game over, all assets in code. Came back: Iron Tide (Opus, stopped at our 150-minute limit) and Duskfall (Fable, 75 minutes).
- Phone finance app. Asked for: a phone-sized app on a supplied dataset with a four-step send flow. Came back: two apps that update the balance correctly, one with haptics (Opus).
- One-lane MOBA. Asked for: a League-style lane with click-to-move, Q W E R, minions, towers and an AI hero. Came back: two lanes both named Riftline, each with a shop; Opus 5.5's was still being refined at our 150-minute limit, Fable 5.1's finished in 131 minutes.
- Open brief. Asked for: invent a product and make its site memorable. Came back: Nod Night Railway and Yonder.
- Kanban board. Asked for: a team board on supplied cards with drag and drop, undo, filters and a details panel. Came back: two compact boards that look strikingly alike, with red-flagged overdue cards, undo and auto-saving details panels; Opus 5.5's also tracks "3 of 14 done".
What do these prompts have in common?
A shape you can reuse. Each one fixes the content the model must not invent, lists the features or flows as bullet points, states technical rules such as "a single HTML file that opens from disk" or "three.js is provided as a local file", and asks for a phone layout where relevant. None of them tells the model what it should look like. That gap is where Claude Code's choices show, and it is why Opus and Fable returned such different results from identical text.
How do you use these examples?
Copy the prompt closest to what you want to build, replace our product, data or game with yours, and keep the structure. Then run it in Claude Code and compare your build with ours. To see what the same prompts produced in OpenAI's agent, open any brief and switch to the GPT builds.
Same short prompt, four very different answers. Pick your favourite before you see who made it.
Take the blind test
Frequently asked questions
Where can I find good Claude Code examples?
Here: eight prompts, each with the exact text Claude Code received and the two builds it produced with Claude Opus 5.5 and Claude Fable 5.1, all live and unedited. Every prompt has a copy button on its brief page.
What makes a good Claude Code prompt?
Fixed content, a list of features, technical constraints and a clear definition of done. Every prompt here names the product and data, lists the sections or interactions, sets rules such as a single HTML file, and requires a phone layout.
How long are these prompts?
Between one short paragraph and about a page. The open brief is the shortest and the most open; the game and app briefs are the longest because they list systems and flows step by step.
Do the examples show Claude Code's best work?
They show its first and only attempt at each brief, which is the point: no cherry-picking, no retries for a better result. The only retries were for infrastructure failures, listed on each build.