Skills: teaching LIA your way of doing things

The problem

You have habits LIA can't guess. Your weekly review always follows the same order, your meeting notes have a precise format, you prefer a weather table to a paragraph. Explaining it again every time is tedious, and the result varies from one time to the next.

What I wanted

I wanted this know-how to be written once, in a format a human can read, and for LIA to use it when it applies. That's a skill: a SKILL.md file, in the open agentskills.io format that other assistants share, with a name, a description and instructions. It can also bring reference documents, templates and scripts. Depending on its type, a skill:

gives expert instructions, for example for writing a professional letter;
describes a method LIA applies with her own tools;
sets a plan template, like a morning briefing that gathers calendar, tasks, weather and e-mails;
runs a small script that produces an interactive widget (a map, a month calendar) or an image (a QR code, a chart). A widget stays attached to its message and reopens after a reload.

How it works

At first, LIA only reads the name and description of your skills. When your request matches one of them, she loads its instructions, then its files if she needs them: a big skill costs nothing as long as it isn't used. A plan template declared as fixed, when it matches your request, runs directly, without LIA having to build a plan, which is faster and costs less. If it includes a step on a service you haven't connected, that step is simply skipped. A badge on the answer tells you which skill was used.

To create one, the simplest way is to ask: “create a skill that shows a QR code”. LIA asks you a few questions, picks the right type, writes and checks the skill, then offers it to you on a card with its name, its description and its files. Nothing is installed without your click. You can also ask her to modify one of your skills: she rereads the whole package and regenerates it entirely, so that the description, scripts and documents stay consistent, then shows you what will be replaced, added and removed.

For a game, a quiz or a calculator you want once, no skill is needed: ask LIA, and the page runs in the conversation.

Scripts, under a bell jar

By default, a skill script runs in a disposable container, created for that call and destroyed right after: no network, no special rights, limited memory and time, and nothing of LIA's inside, neither her code nor your credentials. If that container can't start, the script doesn't run: there's no fallback to a less protected mode. A command that needs to download a package names the sites it wants to reach; it then goes through a single intermediary, to those sites only, once you or the administrator have authorized them. The files it produces come back to you in your gallery, after a check that their content really matches their type.

Skills written by others

From the settings, you can search the public skills.sh library or read a GitHub address. You read a skill before installing it: its source, its exact version, each of its files and what the published audits say. A skill judged too risky is refused, and what gets installed is exactly what you read.

A skill from the library, from a web address or from a plugin was written elsewhere, and LIA treats it that way. Its instructions run separately, with its own files and those you attach, without access to your mailbox, your calendar or your connectors. It displays no image loaded from the web and no widget, and triggers no plan template. A skill you import from your own file, on the other hand, works with your connected data: only keep the ones whose source you know.

The limits

There's no version history: a replaced skill can't be recovered, hence the card that details the changes before your click.
Skills shipped with LIA and those from a library or a plugin can't be modified by chatting: the former are maintained by the administrator, the latter follow their source.
The number and size of skills are capped by your instance.