Skills

Two kinds

A prompt skill changes how an agent behaves. A code skill does work the agent cannot do with the connectors it already has: parse a file, call an API, calculate a date. They are attached in different places and they are not interchangeable.

Code skills are next. Everything after that is prompt skills.

Code skills

A code skill is a Python function you write, stored in your workspace, and published as its own isolated function. Once it is attached to an agent, the agent can call it like any other tool. The function must define main(input, context). input is the arguments the agent passed. context includes the organisation, agent, and session ids.

• Timeout is 1–30 seconds. The default is 10.
• Source, input, and output are each capped at 64 KB. The agent sees at most 16 KB of the result.
• An agent can call code skills at most 25 times in one run.
• Network access is off unless you turn it on. With it off, the function cannot open connections.
• Secrets are write-only. You can replace them, but the saved values are never shown again.
• Code skills are available on Pro (5 skills) and Business (25 skills). A workspace can start at most 30 deploys an hour.
• Every run of a code skill is metered as “Code skill runs” on Plan & Usage, including runs that fail.

Prompt skills

A prompt skill is a short prompt fragment stored in your workspace's Skills library. When you attach one to an agent, its text is injected into that agent's system prompt on every run — you do not need to copy anything into the instructions yourself.

Skills are meant to be narrow and reusable. A skill called “Silence when idle” does one thing: it tells the agent to end the run quietly when there is nothing to act on. The same skill can be attached to every agent in your workspace, saving you from writing that rule into each one individually.

Default skills

Every new workspace starts with eleven default skills across four categories:

• Behaviour — Silence when idle, No duplicate actions, One action per run, Urgency first, After-hours notifications only, Escalate when unsure
• Output — Professional tone, Concise run summary
• Safety — No PII in summaries, Stop gracefully at budget
• Domain — Lead qualification

Default skills are fully editable. If “professional tone” is not the right voice for your business, edit the content to match what you actually want. Editing a skill updates it for every agent that has it attached.

Attaching skills to an agent

Open an agent, go to the Agent tab, and scroll down to the Skills section. Every skill in your library appears as a pill. Click a pill to attach it (it turns green); click again to detach. Changes take effect on the next run — no publish required.

Injection order

When an agent runs, its system prompt is assembled in this order:

• Platform safety preamble (always present, not editable)
• Agent instructions (the text from the Instructions box)
• Attached skills (each skill's content, in alphabetical order by category then name)
• Reasoning preamble (if the Reasoning setting is on)
• Episodic memory recall (if enabled)

Skills are injected after the agent's own instructions, so the instructions take precedence if they conflict with a skill. In practice, skills are designed to be additive — they cover concerns the instructions typically omit.

Managing your library

Go to Skills in the sidebar to see your prompt-skill library. From there you can:

• Preview the content of any skill by clicking Preview
• Edit the name, tagline, category, or content of any skill
• Create a new skill with your own prompt fragment
• Delete a skill — agents that had it attached will simply stop receiving it

Writing a good prompt skill

• One concern per skill. A skill that tries to handle tone, safety, and output format is harder to reuse and harder to reason about.
• Use imperative commands. “Do not send more than one message per run” is clearer than “you should try to limit outbound messages.”
• Keep it under 150 words. Longer fragments are harder for the model to follow reliably and use more tokens on every run.
• Avoid duplicating connector logic. A prompt skill cannot call tools. It only tells the agent how to behave. Running your own code is a code skill, above.