0.21.0

Skills is now called Knowledge. The section moved to a new address and every link you had still works. Inside it, collections and skills are unchanged in name — what changed is nearly everything about working with them.

AI clients can no longer write skills

0.16.0 said a profile could be allowed to write to a collection. That is gone. An attached collection now exposes read-only tools: an AI client can follow your procedures and can no longer create, change or publish them. Writing happens in the dashboard.

Nothing you have set up breaks — no profile in production had been granted those write tools, so there is nothing to reconnect or re-tick. We are calling it out because 0.16.0 promised that if the shape of Skills changed we would say so in a release note rather than change it under you.

The reason is simple. A skill is an instruction your AI client follows. Letting the same session both follow instructions and rewrite them means anything that can talk to that session can change what it does next. Authoring moves behind the dashboard, where a person does it.

A skill is a folder, and you edit it where you read it

Opening a skill now shows its contents as a file list: SKILL.md plus the references, scripts and assets folders. Expand any document to read it. Switch it from Preview to Markdown and you are editing, in place — there is no separate editor screen to go to and come back from.

You can add files by uploading them or creating them, rename them, and remove them, all from the same list.

One save for the whole skill

Nothing is written until you press Save. Editing the body, adding two reference files and renaming a third is one save and one entry in the history, rather than five. Every unsaved change is named on the page and listed again before it is written, a removed file stays on screen with an undo until you save, and leaving the page with unsaved work asks first.

Publishing is blocked while anything is unsaved, so what you publish is what is on your screen.

Every change is recorded, and you can put any of it back

History covers the whole skill, not just its body. Each entry shows which documents changed, who changed them and when, and a line-by-line comparison of what that save actually did. You can read any past version in full and restore it — which writes it as a new version rather than erasing what came after, so the record stays intact.

This matters more here than anywhere else in FloConnector: a skill is an instruction, so a change to one changes how your AI client behaves on its next call. Now you can see exactly who changed what, and undo it.

Longer skills

A skill can now be 20KB rather than 8KB. The old limit was rejecting most real-world skills on import — of the published Agent Skills we tested against, well over half were refused purely for length. 20KB matches the format's own guidance. Long detail still belongs in a reference file, which is only loaded when the skill points at it.

Bringing skills in from elsewhere

Import now opens with a drop zone and a packaging guide you can download and hand to your AI client. It sets out the folder layout, the frontmatter and the size limits, so the model can build a correctly-structured bundle for you rather than you assembling one by hand.

Collections get a mark

A collection can carry an emoji, picked when you create or edit it, and the card takes its colour from it. It is display only — your AI clients never see it.

Still in alpha

Knowledge remains in alpha. This release is the reason that label exists: the structure did change, and we said so here first.