Custom Skills

Custom skills extend what agents can do beyond the built-in set. A skill is a plain markdown instruction block paired with an allowedTools list — when assigned to an agent, the content is injected into its system prompt and the tool permissions are granted.

Creating a skill from the UI

1. Open Settings → Manage Skills.
2. Click + New Skill.
3. Give it a unique slug (e.g., my-deploy-helper), a name, and a description.
4. Write the skill body in the Content field — plain markdown that the agent will read as instructions.
5. Set the allowedTools list — one entry per line (e.g., Bash(curl:*), Read, Edit).
6. Optionally set a model override or context mode.
7. Save, then assign to a class or individual agent.

Skill body

The body is injected verbatim into the agent’s system prompt under a named section heading. Write it as a direct briefing to the agent:

# Deployment Helper

Use this skill when deploying to staging or production.

## Steps
1. Run `npm run build` first and verify exit code 0.
2. Use `{{DEPLOY_TOKEN}}` for authentication — never hardcode it.
3. Always notify the team via Slack after a successful deploy:
   `curl -X POST {{SLACK_WEBHOOK}} -d '{"text":"Deployed ✓"}'`

## Allowed commands
Only run deployment commands explicitly listed above.

Use secrets in skills

Skill content supports {{SECRET_KEY}} placeholders. They are resolved before injection — the agent sees the resolved value, but it never appears in conversation history.

Tool permission format

The allowedTools list controls which tools the agent can call when this skill is active. Each entry maps to a Claude Code tool name, optionally with a command filter:

Entry	What it allows
Read	File reads
Edit	File edits
Bash	Any bash command
Bash(git:*)	Only git … commands
Bash(curl:*)	Only curl … commands
Bash(npm run:*)	Only npm run … commands

Multiple entries are combined — an agent with Bash(git:*) and Bash(curl:*) can run git and curl, but not arbitrary shell commands.

Context mode

Skills support two context modes:

• inline (default) — instructions run in the agent’s main context. All turns share the same conversation.
• fork — the skill spins up an isolated sub-agent for its work and returns a result. Useful for expensive operations you don’t want polluting the main context.

Hot reload

Updating a skill’s body triggers an automatic restart for all agents currently using that skill. The CLI session is preserved via --resume, so the agent picks up the updated instructions on its next turn without losing conversation history.

Token cost

Every skill adds tokens to every agent turn that uses it. A 500-word skill body on an agent with ten skills adds ~5 000 tokens of overhead per turn. Keep skill bodies concise and avoid repeating instructions already in the global System Prompt or class instructions.

Registering skills from code

Built-in skills are TypeScript files in src/packages/server/data/builtin-skills/. Each exports a BuiltinSkillDefinition object:

export const mySkill: BuiltinSkillDefinition = {
  slug: 'my-skill',
  name: 'My Skill',
  description: 'What this skill does',
  content: '# My Skill\n\nInstructions here…',
  allowedTools: ['Bash(curl:*)'],
  assignedToAll: false,
};

Add the export to the skills registry file to make it available in the UI. This is the right approach for skills that should ship with Tide Commander itself rather than per-user customisation.