Free Claude skill

BounceTest.

Seven Claudes get five seconds on the top of your page and say what you do andwhether they stay or leave.

It is a simulation with seven personas, not real visitor data. It measures nothing about speed, SEO or conversion.

Free with an account. No password: you get a login link by email, and it stays in your library.

Price
Free
Made by
Pontus Henriksson
Works in
Claude Code and Claude.ai
What it is
One SKILL.md file, zipped in a bounce-test folder

How it works.

  1. 01

    Give it the page

    A URL if the tool can fetch it, or a screenshot of the top of the page plus its text, or the text only, in which case the phone verdict is marked as a guess. If the fetch fails or shows a cookie wall, it asks for a screenshot. It also asks what the company sells, who the ideal customer is and what a visitor should do first. Then it lists the first screen, element by element, and tests exactly that.

  2. 02

    Seven visitors, five seconds

    Each visitor sees only the first screen, has not scrolled and knows nothing else about the company. Each says what it thinks the company does, who it is for, whether it stays or leaves and why, and one concrete change that would keep it. None of them sees another's answer.

    • 01The buyer in a hurryReady to buy today. Wants to see the offer and the next step at once.
    • 02The phone userThumb on the screen. Judges what fits and what can be tapped.
    • 03The skepticHas been burned before. Looks for proof: real names, reviews, clear terms.
    • 04The price hunterWants to know roughly what it costs, or who it is priced for.
    • 05The strangerNever heard of you. Leaves when jargon replaces a plain sentence.
    • 06The competitorRuns a similar business. Says what is generic and what is different.
    • 07Your target customerYour ideal customer. Stays if the page clearly speaks to them.
  3. 03

    The verdict

    It counts how many understood the offer (and how many partly), how many stayed, and the main reason people left, quoted from a visitor. Then it names the one element that cost the most visitors: the headline, the subheading, the button, the image or the clutter.

  4. 04

    Five new tops

    Five new versions of the top of the page, each with a different angle: the outcome, the customer named, the problem, the proof and the plain description. A headline of at most ten words, a subheading of at most 20 and a button of at most four. Only your facts are used, and a version that needs facts you have not given is marked "needs your facts" and left out of the vote. The visitors vote blind, and you get the votes and a recommended winner. You can pick another.

  5. 05

    The fix list

    At most six fixes, most important first. The first is always for the element that cost the most visitors, and each is tied to the visitor who raised it, with a phone view fix if the phone user left. Then it asks whether to run the seven visitors again on the winning version, one more round at most.

What you get back.

Seven short visitor verdicts, a count, five new tops with votes and a fix list. This is the shape the verdicts and the count come in.

The formatNot a real result

Each visitor, at most 80 words

They do
What they think the company does.
For
Who they think it is for.
Verdict
Stay or leave, and why.
Stay if
One concrete change.

The count

Understood
_ of 7, and how many partly.
Stayed
_ of 7.
Main reason
The reason given most often, quoted from a visitor.
Answers in your language. The new versions are written in the page's language.

It never claims what it did not measure.

No page speed, SEO scores, bounce rates or conversion rates, ever. If you ask, it says which tool would measure them, for example PageSpeed Insights or your own analytics. The new versions never invent prices, numbers, results, reviews or guarantees: it writes a placeholder and asks.

[ADD: the result you got]

It is honest, not kind: a vague headline gets seven leave verdicts. It never asks for passwords, logins, card numbers or ID numbers. For a page behind a login, you give a screenshot with personal data and customer details blacked out. It is not for full site audits or writing the whole page.

Install.

Claude Code

  1. 1Download the zip and unzip it.
  2. 2Put the bounce-test folder in ~/.claude/skills/ for every project, or in .claude/skills/ inside one project.
  3. 3Say "test my website" or type /bounce-test, with a URL or a screenshot of the top of the page.
unzip ~/Downloads/bounce-test.zip -d ~/.claude/skills/

Best here. The seven visitors run as separate agents, so none of them sees another's answer.

Claude.ai

  1. 1Download the zip. Keep it zipped. Code execution must be on (Settings › Capabilities).
  2. 2Go to Customize › Skills, click +, then upload the zip.
  3. 3Paste a screenshot of the top of the page and its text, or just the text, and ask Claude to bounce test it.

There are no agents here, so Claude plays the visitors one at a time. The blind test is simulated in one conversation.

See it the way they do.

Real visitors are quicker and less polite than seven personas. For speed, SEO and conversion, use a tool that measures them.