# Browser recipes

> What your agents learn to do on each website, so they can repeat it in seconds, and what they save from the pages they study.

Skills, memory and library · Updated: 2026-10-16 · version 2.0.145 · https://nobreiniers.com/en/help/recipes/

When an agent does something on a website, it learns how. That recipe lets it repeat the job much faster later.

## See the recipes

1. In the sidebar, press «Recipes».
2. Each recipe says which site it works on and what it does: how many steps it has, how many times it was repeated and how many went fine, and whether it had to repair itself.
3. To find one, type in «Search recipes» or filter by «Verified», «Drafts», «Outdated» or «Archived».

On a recipe you can:

- Press «Trust» so the agent repeats it without asking you first, and «Stop trusting» to have it ask again. A step that pays or creates a new account is listed under «Asks you before»; once trusted, the list says «Trusted: it does not ask you before». If the recipe later learns a new step that pays or creates an account, it is no longer trusted and asks you again.
- Press «Archive» to stop using it (it stays kept) and «Restore» to use it again.

A value that changes each time shows under «What changes each time». A password or code shows as «Secret: asked each time, never kept».

## Teach a recipe

1. Ask an agent for a task on a site, for example «post this on my page».
2. If the agent needs your help, do it yourself in the app's browser: the app records what you do so the agent learns it (passwords and codes are never kept).
3. Next time, the agent repeats the recipe by itself.

A task teaches a recipe whatever the pages it read said: what a page says is never a reason to keep one from being learned.

> **Note:** A recipe repeats a job. What tells an agent where things are on a site (menus, buttons, search, sign-in) is its site map, which it builds by itself. See [the app's browser](https://nobreiniers.com/en/help/browser/).

## What stops a recipe

A recipe runs from start to end without asking you, in every permission level. Following, joining, liking, posting or sending is the work you asked for: what it types is checked where it leaves, as when the agent types it. It stops only before:

- a step that pays or creates a new account: it asks as that same click would live (paying always asks);
- a secret, such as a password or a code: asked for each time and never saved;
- a button or a field one of your own rules may cover: your rule decides.

A value that has to be read off the page each time is never guessed: the agent types it itself at that step.

A button's words no longer decide anything, nor do the calls a page makes by itself (counters, logs).

## When a site changes its buttons

If the site renamed a button since the agent learned the recipe, the recipe heals itself and goes on:

- One added word is accepted: «Join» that now reads «Join group» is still that step.
- If the button already shows the state the step leads to (for example «Following» instead of «Follow»), the recipe takes it as done and does not press it again: an account you already follow is never unfollowed.
- It never heals into the opposite («Follow» → «Unfollow»), into a payment, a new account or a secret, or into another kind of action.
- If the site goes through a sign-in or another address before it arrives, the recipe waits and acts only on the recipe's site.
- A button the page redrew is found again by what it is and its name, and an outcome already on the page counts as done.

When you ask for something on a site that has recipes, the agent knows them from the start of the turn, before it opens the site.

## A recipe and its skill go together

When an agent learns how to do a job on a site, what it learned is saved in one pass as a skill and its recipe, tied to each other: the skill says the steps in words and the recipe is its part that runs on the web. Next time the agent picks the skill and replays the recipe in one go, with the values of this time (the person, the text) and not the ones it saw before. The names of your own accounts, pages and groups are never kept in them.

## A recipe is checked before it is trusted

A recipe that only reads (it opens, looks, copies) is replayed once in the background, in a tab of its own, to check that it still works, before the agent leans on it. A recipe that could change something (post, send, follow) is not tried for its own sake: it counts as unverified until its first real use. A recipe that stops working and is not relearned within two weeks is set aside by itself.

## What an agent learns on a page is saved at once

While it studies a page, an agent may note something worth remembering: a fact about the site, or something about how you work. Since 2.0.141 that note is saved to its memory right away, and you have nothing to do.

There is no exception: whatever the page said, the note is saved at once, and so is what you say yourself. See [Memory](https://nobreiniers.com/en/help/memory/).
