Skills

Skills are reusable, named AI tools that can be invoked on demand. Unlike agents (which are personas), skills are task-specific tools that produce structured output with automatic secret injection.

Creating Skills

Field	Description	Example
`name`	Unique identifier	`code-reviewer`
`description`	One-line summary	"Review code for quality and security"
`instructions`	Full system prompt	Markdown with role, rules, output format

Field	Description
`tags`	Filtering tags (e.g., `security`, `review`)
`model`	LLM model override
`agent`	Agent whose prompt is prepended
`requiresEnv`	Required environment variables
`envMapping`	Map runtime vars to vault secrets
`requiresBins`	Required binaries on PATH

Example

Create skill "pr-summary":
  Description: Generate a pull request summary from a diff
  Instructions:
    You analyze git diffs and generate concise PR summaries.
    Output format:
    1. Title (max 72 chars)
    2. Summary (2-3 sentences)
    3. Changes list (bullet points)
    4. Testing notes

Secret Injection

Skills can securely access vault secrets at runtime via environment mapping:

Create skill "deploy-notify":
  Required env: SLACK_WEBHOOK_URL
  Env mapping: SLACK_WEBHOOK_URL → SLACK_DEPLOY_WEBHOOK
  Instructions: Send deployment notifications to Slack
    using process.env.SLACK_WEBHOOK_URL

At runtime, the vault decrypts the secret and injects it as an environment variable. Secrets are never exposed in logs or responses.

Testing

Write Instructions Define the skill's system prompt with role, rules,

and output format.

Test in Sandbox ``` Test skill with instructions: "You are a commit

message generator..." Message: "Added user authentication with JWT" ``` Testing runs in isolation — nothing is saved.

Save When Satisfied The skill becomes immediately available via MCP

tools, chat, and pipelines.

Invoking Skills

💬 Chat

"Invoke the code-reviewer skill on the auth middleware module."

🔌 IDE (MCP)

The invoke_skill tool is available in your IDE's AI context.

🤖 Agents

Agents can call skills during execution for specialized sub-tasks.

🔄 Pipelines

Skills can be nodes in automation pipelines (e.g., PR review → summary → comment).

Management

Action	MCP Tool
List all	`list_skills`
View details	`get_skill`
Update	`update_skill`
Delete	`delete_skill`
Reload from disk	`reload_skills`

💬 Chat

🔌 IDE (MCP)

🤖 Agents

🔄 Pipelines

On this page