Skills
Every skill is registered in the ConvexskillDefinitions table and can be attached to any agent definition.
| Field | Description |
|---|---|
skillId | Unique identifier |
name | Human-readable skill name |
content | The SKILL.md body — conventions, procedures, examples |
source | Origin: custom, template, or marketplace |
version | Semantic version string |
status | Quality pipeline stage (see below) |
tags | Discovery and categorization tags |
owner | Org or user who owns this skill |
Quality Pipeline
Skills progress through a review pipeline before reaching production use:draft → in-review → approved → active. A
Skills are pattern-matched at runtime. The agent receives a skill directory and decides which skill to activate based on task context. This is an LLM decision, not hard-coded routing.
contentHash field detects post-review drift — if skill content changes after approval, the hash mismatch signals re-review is needed.Tools
Tool bindings give agents access to external capabilities: APIs, databases, file systems, shell execution, web search, and MCP servers. Agents declare tools in two categories:- Required tools — always available, cannot be removed at deployment. Define the minimum capability surface.
- Optional tools — can be added at deployment time. Expand the surface for specific use cases.
toolRecommendations field. When a skill is attached, its recommended tools inform which optional tools should be enabled — connecting “what the agent knows” to “what the agent can do.”
In Burgundy: Manage skills from the resources page. →