SKILLS
Your methods, written once, followed every time.
A skill is a record: a title, one line saying when to use it, the method in plain words, and the files that go with it.
Not code, an instruction
A skill is written the way you would explain a procedure to a new hire. An overview, when to use it, the instructions, some examples, and what to avoid at all costs.
That outline is already in the record when you create it. You fill it in, you do not invent the structure.
No code, no deployment, no ticket. It is a record in your workspace, editable like any other, by the people who actually know the work.
The agent sees the table of contents, not the book
In every conversation the agent carries the index of your skills: the title, and the line saying when to use it. Nothing more, and forty entries at most.
The instruction that comes with that index is unambiguous: before any specialised task, open the skill, and do not improvise.
It then opens the whole record, and only that one. The context stays short, the method stays complete.
The attachments follow
A skill can carry files: a letter template, a rate card, an example of a deliverable that went down well.
When the agent opens the skill, those files are copied into its working environment, into a folder named after the skill. It does not look for them, it has them.
What your teams write cannot break everything
Four skills ship with the platform, among them office document production and driving a browser. Their names are reserved: a record in your workspace cannot take their place.
A skill that requires an execution environment stays invisible to an agent that has none. It cannot promise what it has no means of delivering.
And any text coming from your records is sanitised before it enters the model's context. Control and bidirectional characters, the ones used to hide an instruction inside innocuous-looking text, are stripped out.
How it works
Your firm has its own way of writing a summary note.
You create the record: a short name, and the method as you would dictate it. The expected outline, the tone, the figures that must appear.
You say when to use it, in one line of a hundred and sixty characters at most. That is the line the agent reads permanently, and it decides everything.
You attach the model: the last note that went to committee, the one everybody quotes as the reference.
The agent sees it go by. The title and the « when » line enter its context in every conversation, without anyone switching it on.
It opens it at the right moment. Someone asks for a note, it opens the skill before writing a word, and the attached model lands in its environment.
The other pieces of this chapter
Reflexes
The computation an agent wrote once, replayed identically.
Environments
An isolated, tooled machine that sleeps without losing state.
Hybrid search
Meaning and wording weighed together, filtered in the engine.
Agent memory
Memories that are records, not a black box.
Business documents
Scans become text, deliverables become real files.