Skills

A skill is a self-contained playbook, a procedure, a prompt template, a checklist, that any agent can invoke. Build a skill once, hand it to every agent that needs it.

Tools are what an agent can do (call GitHub, search the web). Skills are what an agent knows how to do, the higher-level know-how that strings tools and prompts together to accomplish a task type. A skill is to an agent what a runbook is to an on-call engineer: a documented procedure to follow when a familiar situation comes up.

Anatomy of a skill

A skill bundles:

• A name and description - what the skill is for; how the agent should recognize when to use it.
• A trigger - keywords, task patterns, or explicit invocation that surface the skill to the agent.
• Steps - the ordered playbook. May include prompts, tool calls, decision branches.
• Inputs - what the skill needs in order to run (a URL, a dataset name, a target audience).
• Outputs - the shape of the deliverable when the skill finishes.

Some skills are tiny (a single, well-tuned prompt). Others are multi-step procedures with tool calls and branches. Both are valid.

Built-in skills

Busy Bee ships with starter skills attached to each role:

• PM: "Scope a feature," "Write acceptance criteria," "Break work into tasks."
• Developer: "Open a PR with a generated description," "Investigate a flaky test," "Refactor a file safely."
• Researcher: "Profile a company," "Compare three competitors," "Pull recent news on a topic."
• Marketing: "Draft a launch post in our voice," "Write a 3-step outbound sequence."
• QA: "Write tests for a function," "Reproduce a bug from a description."

Open an agent's settings to see which skills it has access to.

Creating a custom skill

The simplest skill is a saved prompt. From the Skills page:

1. Click + New skill.
2. Name it (e.g. "Draft launch checklist") and describe when it should be used.
3. Paste the prompt template the skill should run.
4. Optionally define inputs (the skill will ask the agent for them at runtime).
5. Save and attach to the agents that should have it.

For more complex skills with multi-step logic, the editor accepts a structured definition, a sequence of steps, each one a prompt or a tool call, with simple branching based on outputs. This is closer to writing a tiny workflow scoped to a single agent.

When to extract a skill

If you find yourself:

• Pasting the same long prompt into the chat input across multiple tasks → extract a skill.
• Telling an agent the same procedural correction more than twice → bake the correction into a skill.
• Wanting consistency across team members using Busy Bee → ship a shared skill rather than relying on each person to phrase requests correctly.

Skills compound. The team gets better at its work, automatically, every time you ship one.

What's next

Skills are know-how. Tools are the levers an agent pulls to act on that know-how.