Skills

A skill is an instruction for one task: how to recognise it, what to do step by step, where the pitfalls are and how to check the result. “How to put together the weekly report”, “how to file an insurance claim”, “how to answer a complaint about delivery”.

The assistant only sees skill names and “when to use”, and opens the full text by itself when a task fits. So you can have lots of skills — each conversation gets barely more expensive.

The format is the open Agent Skills standard (a SKILL.md file), the same one used by skills for Anthropic, Hermes and OpenClaw. So you can simply give a ready-made skill from the internet as a link.

How to teach it

Say	And it will
“Remember how you did that and always do it this way” — to the assistant itself, in the linked chat or the dashboard	draft a skill from what it just did and tell you where to save it. It only starts following it after you tap “Save”
“No, do it like this: …” — to the assistant when it got something wrong	likewise offer to remember the right way
“Remember a skill: how to put together the weekly report — …” — to the builder	ask about anything missing and save the skill right away
“Here's a skill from the internet — add it: …link…”	read it, retell what it will make the assistant do, and name the risky parts separately. It adds it only after your “yes”
“What skills does my assistant have?”	list the skills and, separately, drafts waiting for your consent
“Save the skill it suggested”	save the draft
“Forget the skill about reports”	remove it; “undo” brings it back

Drafts: “Wants to remember”

After a task the assistant handled for the first time, or after you corrected it, it may suggest a skill. It can't save it by itself: a draft doesn't take effect until you tap “Save” on the panel — the “Skills” line, “Wants to remember” section. “No” — and the draft is gone.

The assistant can suggest skills only in a conversation with you — in the linked messenger chat or in the dashboard. A customer on your site can't teach it anything: an instruction from just anyone that would apply in every following conversation is a way to break the assistant, not teach it.

If the assistant drafted a skill after reading someone else's text — a page, an email, search results — it's marked “Suggested after reading: …”. Such an instruction could have come from there rather than from you, so it can only be saved with a second tap (“Yes, save”) and only below its text: the first tap expands the text itself. If the assistant rewrote the draft while you were reading, it won't be saved — read the new one and approve again. The builder will also ask for consent to this particular draft separately: a “yes” to another skill doesn't save it.

A skill from the internet

Give a link — to the builder in words or in the “Skills” sheet (the field at the bottom, “Read”). These work:

• a skill folder on GitHub — https://github.com/…/tree/main/skill-name;
• the file itself — https://github.com/…/blob/main/…/SKILL.md;
• any link to a raw SKILL.md.

First you'll see the full text and the lines worth pausing on: where something gets sent, which addresses and keys are named, whether money is involved, whether the text refers to scripts, whether it tries to override the assistant's other instructions. This is a hint, not a filter: an empty list doesn't mean “safe”.

Good to know:

• Nothing is saved before consent. The builder asks “add it?” and adds it only after your answer — not in the same turn it read it.
• Exactly what you saw is added. If the file at the link changed between showing and consent, adding fails — read it again.
• Only the SKILL.md text is taken. Scripts, references and other files from the skill folder are never taken and never run. If a skill says “run scripts/…”, the assistant will say it can't do that step.
• A skill doesn't grant permissions. The standard's allowed-tools field has no effect: only the owner gives the assistant permissions. Outside actions still go the way you set them up — with confirmation, if that's on.
• An added skill is someone else's text that the assistant will follow in every conversation. Read it as carefully as you would an instruction for a new employee.

Limits

What	How much
skills at once	30
drafts waiting for consent	10
names and “when to use” of all skills together	8000 characters
text of one skill	20000 characters

The name is lowercase letters, digits and hyphens, no spaces: weekly-report, insurance-claim. The builder or the assistant itself will come up with it.

When there's no room, a new skill isn't added and an old one isn't pushed out: remove what you don't need first.

Undo

Every change to skills — adding, rewriting, removing, saving a draft — goes into the change log, and “undo” reverts it: the previous text, the removed skill, and a saved draft becomes a draft again. A rejected draft doesn't go into the log: it never took effect.

Who can see them

• Every dashboard member sees the list of skills and their texts — they're the assistant's settings, just like its personality.
• An admin can change them — save, remove, add by link.
• Skills are included in the data export (the Account section → “Your data”) and are erased together with the assistant.

Not available yet

• Self-learning without your consent doesn't exist on purpose: the assistant only suggests.
• You can't save a skill with a button in the dashboard — only in words to the builder or from the assistant's own suggestion.
• You can't export a skill back as a SKILL.md file yet; the text is visible in the “Skills” sheet.