In September 2026 we shipped a browser extension that puts Prompt Score, Quick Optimize, and your Keep My Prompts library directly above the message box in Claude and Euria. The idea is simple: the prompt that matters is the one you are about to send, not the one you saved last month. So the extension lets you score it before you press Enter, hands you an improved variant in one click, drops that variant into the composer, and quietly saves the prompts you actually send, with a filter that keeps the "yes, shorter" replies out of your library. On launch day one of our own perfectly ordinary prompts went from 3.0 to 4.8 with that one click, and the answer we got back was the one we would have used. The rest is what it does, what it never touches, and how to install it in two minutes.
Why the pill sits above the composer
Every prompt you write in a chat is a small decision made under time pressure: you have a task, you type the first version that comes to mind, and you press Enter. If the answer is mediocre, you rephrase and try again. Scoring a prompt after the fact teaches you something for next time; scoring it before you send changes the answer you get now.
That is the whole point of putting Keep My Prompts inside the chat. You do not switch tabs to a library, find the right prompt, copy it, come back and paste it. The library comes to you, where you actually write the prompts, and does three things in the order you need them: tells you how good the prompt is, makes it better, and remembers it if it was worth sending.
If you work in Claude Code or Cursor, the MCP integration does the same job from the editor. The extension brings the same idea to everyone who writes prompts in a browser.
What you see
Install the extension, connect it to your account, and a small pill appears above the composer in Claude or Euria, with our logo and two buttons: Score prompt and Save to library. It follows the composer as the page changes. You can drag it by the logo if it covers something you need, and collapse it to the logo alone; both the position and the collapsed state are remembered.
The Keep My Prompts pill above a chat composer, showing the score dot, the Score prompt button already used, the Optimize button unlocked and Save to library, with the six-criteria panel open above it
Write your prompt as usual. Once it has at least ten words, press Score prompt. Nothing is saved: the text stays in the composer, and what comes back is a Prompt Score from 1 to 5 on the six criteria we use everywhere on Keep My Prompts (clarity, context, structure, role, examples, and step-by-step reasoning), plus one concrete tip. The score sits in the pill as a coloured dot; click it for the details.
Want to know how effective your prompts are? Prompt Score analyzes them on 6 criteria.
After the score, and only after it, a third button appears: Optimize. It runs Quick Optimize on the text in the composer and shows the improved variant in a panel with two actions. Replace in composer swaps your text for the variant, so what you send is the better prompt. Save to library stores the variant in your library without sending it. Until you press one of those two, or Enter, nothing has been written anywhere: before you send, the extension is a mirror, not a recorder.
Optimize waits for Score on purpose. The score is the baseline that tells you whether optimising is worth it: when a prompt already scores 4.5 or above, the button tells you so on hover. Spending one of your daily optimisations on a prompt that is already good is a waste, and we would rather save you from it.
A real prompt, with the real numbers
Here is the first draft one of us typed on launch day, in Italian, in the Claude composer: a request to write an email to a school principal asking permission for a two-day class trip to Florence, with dates, estimated costs, chaperones, and the learning goals tied to the art history syllabus. An entirely normal prompt. It scored 3.0. Clarity was fine at 5 and Role was clear, but Examples scored 1: the model had nothing to imitate, so it would have produced a generic letter.
Quick Optimize returned a variant that scored 4.8. What changed was not the request, which stayed identical, but its scaffolding: an explicit structure for the email, a short example of the expected tone, and visible constraints on length and register. Examples went from 1 to 4. We pressed Replace in composer, pressed Enter, and what came back was the letter we would actually have sent.
Two things about this example are worth saying plainly. The numbers come from our own account on launch day, not from a benchmark. And the whole exchange took less time than opening a second tab, finding a saved prompt, copying it and pasting it back.
What happens when you press Enter
The second flow is silent. When you send a message, the extension checks whether it looks like a prompt and, if it does, saves it to your library in a folder called From chats. A small badge appears next to the pill, showing the score and "Saved to your library", and fades out on its own after a few seconds. If you had already scored that exact text with the pill, the badge reuses that score instead of calling the model again.
The techniques you're reading about work. Test your prompts now with Prompt Score and see your score in real time.
"Looks like a prompt" is a deliberate filter. Most messages in a chat are conversation: "yes", "shorter", "and point 3?". Saving them would turn your library into a transcript, and the twenty prompts of the free plan would fill up with noise in three days. So the first message of a conversation is captured only if it has at least twenty words, and follow-ups only if they have at least twenty-five, because that is what a genuinely new instruction usually looks like. Everything shorter is ignored.
Sending the same prompt again does not create a duplicate. The library recognises it and bumps a reuse counter instead. That counter is more useful than it sounds: a prompt you have sent five times is the one worth optimising and versioning, and it now stands out in your library.
What leaves the page, and what never does
An extension that reads what you type in a chat has to be precise about its boundaries, so here they are.
Diagram of the extension's data boundary: on the left the chat page with the composer text highlighted as the only thing read, replies and conversation titles crossed out; in the middle the extension with the per-site consent switch and the pause button; on the right the Keep My Prompts server storing the prompt encrypted
The extension reads the text in the message composer, and only when you press Score, Optimize, or Save, or when you send a message on a site where you have turned capture on. It never reads the assistant's replies, the conversation title, the sidebar, or anything on any other site: it runs only on claude.ai and euria.infomaniak.com, and the permissions it asks for say exactly that.
What reaches our servers is the prompt text, the site it was written on, and a timestamp. Captured prompts are encrypted at rest with the same scheme as prompts you create on the website. Scoring and optimising send the text, and nothing else, to the same AI providers the website uses, which we list in our privacy policy.
Consent is per site. The first time the extension sees a supported chat it asks whether you want capture on for that site, and you can change the answer, or pause everything, from the extension icon at any time. The extension authenticates with a personal key you create in your Keep My Prompts settings. The key lives in your browser's extension storage, never in a cookie or on the page, and you can revoke or regenerate it whenever you like. Captured prompts are ordinary prompts in your library: delete them one by one or all together.
There are also daily caps on scoring and capture, applied per account and set by plan. They are sized so that a runaway script or a stolen key cannot run up a bill in your name, and they sit far above what a person does in a day. You will not hit them; a bot would.
In Settings › Integrations › Browser extension, press Create key, then Copy.
Click the extension icon, paste the key, press Connect. The popup shows your plan and how many optimisations you have left today.
Open Claude or Euria. Accept capture for the site when asked, and write your next prompt.
The extension works with Euria both on euria.infomaniak.com and inside kSuite. If you lose the key, there is nothing to recover: Regenerate, on the same settings page, revokes the old one and shows you a new one once.
What the plans include
Prompt Score is included in every plan, including the free one. Quick Optimize has a daily allowance on the free plan and a larger one on Pro and Ultimate. The free library holds twenty prompts: when it is full, the extension keeps scoring your prompts and tells you plainly, with a link, that saving resumes when you upgrade, rather than failing quietly. Details are on the pricing page.
What comes next
ChatGPT and Gemini are next: the same mechanism, with different selectors for their composers. We are also planning a one-click connection, so you no longer have to copy and paste the key. If there is a chat you write in every day and want us to support, tell us at [email protected].
Start with the prompts you already send
If you write your prompts in Claude or Euria, you already have a library; it is just spread across your chat history. Create your account, install the extension, and let the next week of prompts save themselves. When you want a stronger starting point than your own drafts, our templates library holds ready-made prompts for teachers, accountants, lawyers, clinicians, and more, and every one of them can be scored and optimised the same way. And if you want to understand what the six criteria measure and why, this is how Prompt Score works.
