Skills

Why skills

• Write once, reuse everywhere. Instead of copy-pasting the same prompt into every node, author it as a skill and attach it to as many agents as you like. Editing the skill updates every agent that uses it.
• Keep agents focused. On an Agent Loop, the agent sees only each skill's name and description up front. It loads the full instructions on demand when a task matches — so ten specialized playbooks don't bloat the prompt (or your token bill).
• Team knowledge, not personal prompts. Skills are shared with your workspace by default. Admins and the skill's creator can edit; everyone can use.

Create a skill

• Open Dashboard → Skills and click New skill.
• Name (1–64 characters) — the agent sees this in its skill list.
• Description (up to 1,024 characters) — what it does + when to use it. The agent uses this text to decide whether to load the skill, so include concrete trigger phrases, e.g. “Draft a release announcement from merged PRs. Use when asked to announce a release.”
• Instructions — the full playbook. Markdown supported.
• Sharing — Team (default: everyone in the workspace can use it) or Private (only you). Built-in skills are read-only; export one and import your copy to customize it.
• Saving publishes version 1 immediately. Every later save republishes; use Version history to restore a prior version.

Supporting files

• Upload references, templates, or scripts on the skill's edit page. Agents read them at runtime with the built-in read_skill_file tool when the skill's instructions refer to them.
• Limits: 10 MB per file, 30 MB per skill, 500 files.

Attach skills to a node

• Open an Agent Loop or AI Prompt node's configuration and tick the skills you want under Skills.
• Agent Loop — the system prompt lists each skill's name + description; the agent calls the built-in get_skill tool to load a playbook before performing a matching task (progressive disclosure). Requires a tool-capable model.
• AI Prompt — single-shot, so attached skills' full instructions are added to the system prompt directly. Works with every model.
• Agents always run the skill's live published version — edits take effect on the next run, no workflow changes needed.

Import & export

• Export any skill as an Agent Skills-compatible .zip ({skill-name}/SKILL.md + supporting files) from the Skills page or the editor.
• Import a .zip bundle containing a SKILL.md with name and description frontmatter — the open format round-trips cleanly with Claude and other compatible tools.

Knowledge base on LLM nodes

Alongside skills, Agent Loop and AI Prompt nodes can now use the agent's knowledge base directly:

• Agent Loop — enabling Knowledge Base gives the agent a search_knowledge tool for on-demand semantic search over the agent's documents, mid-conversation.
• AI Prompt — the user prompt is searched against the knowledge base and the most relevant excerpts are injected before the call; sources are returned as knowledgeSources on the output.
• The standalone Agent Knowledge node keeps working exactly as before — this option is additive, and existing workflows are unaffected.

Best practices

• Write descriptions as what it does + when to use it — that text is how the agent decides to load the skill.
• Keep each skill focused on one job. Several small skills beat one giant playbook — the agent loads only what the task needs.
• Put long reference material in supporting files rather than the instructions, and tell the agent (in the instructions) when to read which file.