Skills

A Skill is a named, reusable set of instructions written in Markdown. Attach a skill to an AI agent block and it steers that agent's behavior — a repeatable way to encode tone, procedure, or domain knowledge once and reuse it across workflows.

Workspace skills

Workspace skills are selected by the workflow or attached to a conversation. The workspace Skills view and /api/v1/skills endpoints manage these records. A private skill owned by a Confluye member is different from an external customer's skills described below.

Creating a skill

Skills are authored in a SKILL.md-style Markdown document with lightweight frontmatter:

---
name: Support triage
description: Classify and route inbound support messages
---

You are a support triage assistant. Read the message, classify its urgency,
and choose the correct team...

The frontmatter name and description are parsed out, and the Markdown body becomes the skill's instructions. If no name is given, the first heading is used, and the name is slugified. You can also upload a ZIP containing a SKILL.md (or any .md) to import a skill.

Create and manage skills in the Skills view, or via GET/POST /api/v1/skills and PATCH/DELETE /api/v1/skills/{id}. Each skill stores its name, description, instructions, an enabled flag, and usage stats.

Using skills

An AI agent block selects workspace skills through its skill configuration (skillIds, with skillId / skillName supported for a single selection). The Agent runtime first receives their summaries and loads authorized instructions as needed; CLI blocks receive the selected instruction attachments. Workspace skills are also offered in the Command Center and as chat attachments.

Managing skills

The Skills view supports search across name, description, and instructions, plus enabling and disabling. Confluye tracks a usageCount and lastUsedAt for each skill: after a run completes, any skill referenced by a succeeded step has its usage incremented, so you can see which skills are actually in use.

Skills for an external end user

With end users enabled, open Settings → Integrations → End users and memory → Manage end users and memory, select a customer and use Skills for this end user. An explicit workspace Admin or Owner can create, edit, enable, disable or delete these skills. Workspace Members, Viewers and organization-derived access cannot manage them. Your product's gateway API key and the customer's agent cannot create or edit these instructions.

Each skill has a name, description and Markdown instructions. Enabled instructions are added automatically to that customer's Agent, Claude Code and Codex blocks, alongside the workflow's selected workspace skills. They apply even when the workflow has no connected memory resource. The run's persisted workspace, customer and generation determine the selection; a customer selector in workflow input cannot change it. A queued run reads current skills when it starts; edits affect later node execution only when a new preparation loads them, including resume preparation. A running model or CLI call keeps the instructions it already received.

There is no agency skill level. A customer can have at most 32 skills, with up to 8,000 instruction characters per skill. The complete enabled guidance, including names and descriptions, must fit within 16,000 characters; a save that exceeds this limit is refused. Credentials and session tokens are refused. Instructions do not grant new tools, browser access or permissions.

A stale edit is refused with a conflict. Your draft stays visible; Reload latest skill replaces it with the saved revision. Disabled skills stay available in the owner console and export, but are omitted from new run preparation. Names, descriptions and instructions are encrypted together. Customer export and erasure include all skills: erasing the end user deletes enabled and disabled records, including during a worker upgrade, and reusing the external ID does not restore them. A subject-only erasure cannot identify a contact mentioned in free-form instructions; edit or delete those skills separately, or erase the end user that owns them.

Next steps