Skills
A skill is a house method your agents follow. Learn how one is loaded, why its trigger sentence decides everything, and when to write one.
Automate
A skill is a house method your agents follow when the job matches it. It holds the steps, in order, and what must never be done.
Three fields
| Field | What it holds | Limit |
|---|---|---|
| Identifier | Lower case, digits and hyphens. This is the name your agents quote to open the method. | 64 characters |
| When it applies | The condition that makes an agent reach for it. | 1 024 characters |
| The method | The steps, in order, and what must never be done. | 20 000 characters |
How a skill reaches an agent
In two stages, and this is why the trigger sentence matters more than its size suggests.
On every turn, the agent receives only the list of names, about ten tokens each. The method itself is loaded when the agent opens it, and not before. Your credits pay for the names on every turn, for the body only when it is used.
So write the trigger as a condition, not as a summary. A sentence like use as soon as an invoice is more than fifteen days late tells the agent when to open. A title tells it nothing.
Nothing ships by default
The Skills screen is empty on a new project. Skills come only from what you write on this project.
An agent that quotes a name you never wrote finds nothing, and the run says so in plain words, listing the names declared but not found. That line is the fastest way to spot a typo in an identifier.
A skill also keeps an agent working
An automation is suspended when its agent declares tools, none of them has an executor, and no skill is served to it. One served skill removes that third condition.
The agent then reaches the method rather than stopping. It is not a fix for a missing tool, but it is the difference between a report and nothing at all.
When writing one pays off
- The same steps repeat across several automations. Write them once, quote the name in each instruction.
- An order carries a rule that must never be broken. The method holds it, while instructions get rewritten.
- A run failed because the agent improvised a sequence. Fix the sequence in a method rather than in a longer instruction.
A single instruction used by one automation does not need a skill. Put it in the automation and keep the list short.
What the screen refuses
- An identifier with capitals, spaces or accents. Lower case, digits and hyphens only.
- An identifier already carried by another skill on this project.
- A trigger or a method left empty. Both must be filled in.
- More than 30 writes in a minute.
- A viewer trying to change anything. Only the owner and editors write skills.
Beyond 200 skills, the screen lists the first 200 and says so. Search from the field at the top rather than scrolling.
Reading a skill
The card shows the identifier and the trigger sentence, folded on three lines. A skill saved without a trigger is shown as undocumented, which is a good reason to go back and write one. Expanding a card unfolds the trigger, not the method.
Deleting a method
The confirmation is explicit: your agents can no longer follow this method, and its text cannot be recovered.
Automations that quoted its name keep running. Their runs then report the name as declared but not found, so the loss shows up in the history rather than in silence.