Skip to main content
A skill is a reusable, codified workflow: instructions, and optionally scripts and templates, that an agent follows whenever a matching task comes up. Define the process once instead of explaining it every time.

​
When to use a skill

Use a skill for a repeatable procedure with steps worth getting right consistently:
  • “How we format and publish the weekly report”
  • “Our process for triaging an inbound bug”
  • “The steps to onboard a new customer in our tools”
If you have explained the same multi-step process more than once, make it a skill.

​
What a skill contains

Instructions

The step-by-step workflow the agent follows.

Scripts

Optional helper scripts the workflow runs.

Templates & references

Boilerplate, examples, and detailed docs the steps point to.

​
Skills vs. persona vs. memory

A skill is the playbook for one job; the persona is the agent’s character; memory is its accumulated knowledge.

​
Built-in skills

Every agent ships with four skills pre-installed. webhook-receiver covers the full setup handshake for receiving inbound HTTP events from external services: fast acknowledgment, HMAC signature verification, durable cross-run deduplication, and the two-party provider-console flow (endpoint registration, signing secret, validation ping). It works with any provider, including Stripe, GitHub, Linear, and Intercom. Ask an agent to build a webhook receiver and it loads the skill automatically. link-wallet gives your agents access to human-approved online purchases via Stripe Link. The wallet belongs to you, the owner. You log in once and approve every single spend from your Link app, so no card number ever lands in the sandbox. The agent uses link-cli auth login to surface a URL and verification phrase, you approve in the Link app, and the CLI returns a one-time virtual card for the purchase. Spend limits are set by Link ($500 per transaction, $500 per day). Use link-cli --test during development to verify the flow without a real charge. gmail loads when an agent works through a connected Gmail inbox at any real volume. Fetching email one message per request can turn an inbox sweep into hours of quiet bash-loop time. The skill points the agent at the batch pattern instead (list messages with metadata in one request, page through), so a bulk scan takes minutes. It triggers on its own once the job looks like bulk mail. image-generation loads when an agent generates or edits images: posters, diagrams, remixed photos. The agent sends its request to the provider with a placeholder credential, and the platform stamps the real key on the way out, so no key lands in the sandbox and usage bills to your credit balance. The skill covers OpenAI’s Images API and Google’s Gemini image models (Nano Banana), with working examples for both. None of these skills need installation or configuration. They live in the agent’s repository under skills/ and can be read or customized like any other skill.

​
How skills get used

When a task matches a skill’s description, the agent loads that skill and follows it. To teach a new skill, describe the workflow in conversation and ask the agent to save it. You can also give it reference material for the skill to draw on.

Automate it on a schedule

Pair a skill with a recurring run to automate a whole workflow.

Skills live in the agent repository

Each skill is a folder in the agent’s git repository. Clone it to read, edit, or add skills directly.