AI Website Builder — score your prompt before you generate the site

An AI website builder turns one text prompt into a whole page. Every property of that page the prompt does not state, the model decides for you — and it decides in favour of whatever is most common in its training data, which is why underspecified prompts converge on the same blue-and-white landing page with a hero, three feature cards and a contact form.

The published prompting guidance from Anthropic, OpenAI and Google agrees on the fix and words it almost identically: state the goal and the audience, give the structure, supply the real content, name the constraints, and include an example. The scorer above is that advice made countable. It splits a website brief into 9 dimensions and gives each one a weight reflecting how much of the finished page it controls:

Each dimension is graded on a four-step ladder rather than a sliding scale, because the useful distinction is not “how long is this answer” but “is this a label or a specification”. An unanswered dimension scores 0. A few words — enough to name the thing but not to pin it down — scores 0.4. An answer past that dimension’s stub threshold scores 0.7. An answer at or beyond its full threshold (the last column above) scores 1. The thresholds differ per dimension: a call to action is complete in 25 characters, whereas the real content of a site is not.

The score is then the weighted sum: score = Σ weight × coverage. The same figure is computed a second way, from the missing side of the ledger — 100 − Σ weight × (1 − coverage) — and the two must agree to the last decimal. That is what the “cross-checked” badge on the calculator reports, and it is also what powers the gap list: each entry shows exactly how many points writing that one answer would earn.

Alongside the score sits a plainer number. Every dimension has a count of concrete decisions attached — page structure carries five, real content carries six — and the delegated total is Σ ⌈decisions × (1 − coverage)⌉. A blank brief hands the model all 29. That is the number worth watching, because it is the count of choices you will later be editing back out of the generated site.

Finally, the composed brief is capped so it fits a real prompt field. Each answer is cut at 320 characters and the whole brief is measured against 4,000 characters, the limit of the builder at /build. Nine answers at the cap still come in under it, so a brief written here always pastes cleanly.

Bands are matched inclusive at the bottom and exclusive at the top, so a score of exactly 70.0 is “Well specified” and exactly 85.0 is “Build-ready”.

• Anthropic — Be clear, direct, and detailed
• Anthropic — Use examples (multishot prompting)
• Anthropic — Prompt engineering overview
• Anthropic — Token counting (the character-per-token estimate is an approximation)
• OpenAI — Prompt engineering guide
• Google Workspace — Gemini prompting guide

The dimension weights are my own reading of the guidance above, not a figure published by any of those companies — they are a way to make “be specific” countable, and the model is documented in full in the “How it works” section so you can disagree with a weight and adjust for it. The prompt-field limit of 4,000characters is read from this site’s own builder. All sources last checked on 2026-09-05.