Skills
Reusable agent instruction packs — global and project scope, Allowed vs Include in prompt, and MCP discovery.
Skills
Skills are reusable instruction packs stored in ArxDeck (not in your repository by default). Agents discover allowed skills via MCP and can load full content on demand. Skills teach agents how to perform specialized tasks — project conventions, feature workflows, or integration patterns.
ArxDeck Helper can propose skill configuration through project chat. Describe the instructions you want agents to follow and approve proposals in the Proposals inbox. See ArxDeck Helper.
What a skill contains
Each skill has:
- Id — canonical identifier agents and MCP tools use
- Slug — auto-generated from the name; used for repo inject folders and markdown resource links
- Name and short description — shown in skill lists and the discovery block
- Long description — markdown returned by
get_skill(or inlined when Include in prompt is enabled) - Resources (optional) — additional markdown bodies linked from the long description and loaded via
get_skill_resource
Global vs project scope
| Scope | Who manages | Availability |
|---|---|---|
| Global | Platform superadmin | Delivered through content modules; available when an active module installation provides the skill |
| Project | Project admin | Only in the owning project |
Which skills an agent may use is controlled by agent configuration, not by project enablement alone.
Agent skill modes
On each agent's settings, use the Add skill picker to attach skills. Each attached skill has one of two modes:
| Mode | Config field | Behavior |
|---|---|---|
| Allowed | allowedSkillIds (minus injected) | Slug and short description appear in a discovery block in the launch prompt. Agent loads full content on demand via get_skill. |
| Include in prompt | injectedSkillIds | Title, slug, and long description are inlined in the launch prompt before MCP instructions. Must also be in the allowed list. |
The launch prompt discovery block header is: Skills you can load with get_skill when they seem relevant to the user's request:
Include in prompt ids must be a subset of Allowed ids. Global agents may only reference global skill ids. Archived or missing skill ids are removed from both lists automatically on save.
MCP discovery
During a run, agents with MCP enabled call:
| Tool | Purpose |
|---|---|
list_skills | Allowed skills (id, name, shortDescription; optional slug for repo context) |
get_skill | Full skill markdown and resource index (skillId — id or slug) |
get_skill_resource | One resource body by skillId + resourceSlug |
Access requires MCP enabled on the agent, the tool in the MCP allowlist, and the skill id in the agent's allowed list.
Inject in repo (optional)
Project admins can push a project or global skill into the linked GitHub repository:
- UI: Inject in repo on skill rows at Settings → Skills
- Writes
.claude/skills/{slug}/withSKILL.mdand resource files - Requires a linked GitHub repo and write access via organization credentials
This is optional — product skills normally live in the database and are edited in the dashboard.
Configuration MCP
list_skills_config and get_skill_config share the configuration catalog for inspection (global skills plus project-readable skills). Lookup accepts id or slug interchangeably, including stable seeded ids such as skill_setup_helper. Runtime get_skill still requires module installation and agent allowlisting.
Setup & configuration
| Scope | List route |
|---|---|
| Platform (global) | Platform → Skills (superadmin) |
| Project | Settings → Skills |
Typical setup:
- Create or install skills your agents need (name only at create).
- On each agent, use Add skill and choose Allowed or Include in prompt per skill.
- Verify agents call
list_skillsduring runs when instructions should be loaded on demand.