Skip to main content
Two things sit behind the assistant that you never have to ask for. It reaches for them itself, and knowing they exist explains why an answer sometimes arrives in a particular shape.

​
Skills

A skill is a written procedure for one kind of work. Only a one line index of every skill sits in the prompt at all times. When a request matches one, the assistant loads the full body and follows it. That is why the prompt stays small and the behaviour still gets specific. Each also carries a “not for” line, which is what stops it firing on work it does not suit. Asking for a quick explanation of code does not pull in the PDF skill. You do not turn these on. Describing what you want is enough.

​
Installing a skill

You can point the assistant at a public GitHub repository holding skills and ask it to install one.
It fetches that skill’s SKILL.md and any markdown reference files and loads them into the conversation. The repository has to be given as owner/repo, and a branch or tag can be added if you need one other than the default.
Only markdown is fetched. Scripts and other files in the skill are skipped and never run. A skill that depends on running its own scripts will load its instructions and not much else.
An installed skill applies to the conversation it was installed in.

​
Knowledge

Knowledge is reference material rather than procedure. There are currently 475 entries, and the assistant searches them by keyword when a question touches one. The difference is worth holding onto. A skill tells it how to do something. A knowledge entry tells it what is true about something. They are separate mechanisms and neither replaces the other. Knowledge is ours and it is the same for everyone. The things that are yours are memories, indexed sites and custom agents.