Skills Platform

Skills are Triggerfish's primary extensibility mechanism. A skill is a folder containing a SKILL.md file -- instructions and metadata that give the agent new capabilities without requiring you to write a plugin or build custom code.

Skills are how the agent learns to do new things: check your calendar, prepare morning briefings, triage GitHub issues, draft weekly summaries. They can be installed from a marketplace, written by hand, or authored by the agent itself.

What Is a Skill?

A skill is a folder with a SKILL.md file at its root. The file contains YAML frontmatter (metadata) and markdown body (instructions for the agent). Optional supporting files -- scripts, templates, configuration -- can live alongside it.

morning-briefing/
  SKILL.md
  briefing.ts        # Optional supporting code
  template.md        # Optional template

The SKILL.md frontmatter declares what the skill does, what it needs, and what security constraints apply:

---
name: morning-briefing
description: Prepare a daily morning briefing with calendar, email, and weather
version: 1.0.0
category: productivity
tags:
  - calendar
  - email
  - daily
triggers:
  - cron: "0 7 * * *"
metadata:
  triggerfish:
    classification_ceiling: INTERNAL
    requires_tools:
      - browser
      - exec
    network_domains:
      - api.openweathermap.org
      - www.googleapis.com
---

## Instructions

When triggered (daily at 7 AM) or invoked by the user:

1. Fetch today's calendar events from Google Calendar
2. Summarize unread emails from the last 12 hours
3. Get the weather forecast for the user's location
4. Compile a concise briefing and deliver it to the configured channel

Format the briefing with sections for Calendar, Email, and Weather.
Keep it scannable -- bullet points, not paragraphs.

Frontmatter Fields

Field	Required	Description
name	Yes	Unique skill identifier
description	Yes	Human-readable description of what the skill does
version	Yes	Semantic version
category	No	Grouping category (productivity, development, communication, etc.)
tags	No	Searchable tags for discovery
triggers	No	Automatic invocation rules (cron schedules, event patterns)
metadata.triggerfish.classification_ceiling	No	Maximum taint level this skill can reach
metadata.triggerfish.requires_tools	No	Tools the skill depends on (browser, exec, etc.)
metadata.triggerfish.network_domains	No	Allowed network endpoints for the skill

Skill Types

Triggerfish supports three types of skills, with a clear priority order when names conflict.

Bundled Skills

Ship with Triggerfish in the skills/bundled/ directory. Maintained by the project. Always available.

Triggerfish includes five foundational skills that make the agent self-sufficient from day one:

Skill	Description
tdd	Test-Driven Development methodology for Deno 2.x. Red-green-refactor cycle, Deno.test() patterns, @std/assert usage, Result type testing, test helpers.
mastering-typescript	TypeScript patterns for Deno and Triggerfish. Strict mode, Result<T, E>, branded types, factory functions, immutable interfaces, mod.ts barrels.
mastering-python	Python patterns for Pyodide WASM plugins. Standard library alternatives to native packages, SDK usage, async patterns, classification rules.
skill-builder	How to author new skills. SKILL.md format, frontmatter fields, classification ceilings, self-authoring workflow, security scanning.
integration-builder	How to build Triggerfish integrations. All six patterns: channel adapters, LLM providers, MCP servers, storage providers, exec tools, and plugins.

These are the bootstrap skills -- the agent uses them to extend itself. The skill-builder teaches the agent how to create new skills, and the integration-builder teaches it how to build new adapters and providers.

See Building Skills for a hands-on guide to creating your own.

Managed Skills

Installed from The Reef (the community skill marketplace). Downloaded and stored in ~/.triggerfish/skills/.

triggerfish skill install google-cal
triggerfish skill install github-triage

Workspace Skills

Created by the user or authored by the agent in the exec environment. Stored in the agent's workspace at ~/.triggerfish/workspace/<agent-id>/skills/.

Workspace skills take the highest priority. If you create a skill with the same name as a bundled or managed skill, your version takes precedence.

Priority:  Workspace  >  Managed  >  Bundled

TIP

This priority order means you can always override a bundled or marketplace skill with your own version. Your customizations are never overwritten by updates.

Skill Discovery and Loading

When the agent starts or when skills change, Triggerfish runs a skill discovery process:

1. Scanner -- Finds all installed skills across bundled, managed, and workspace directories
2. Loader -- Reads SKILL.md frontmatter and validates metadata
3. Resolver -- Resolves naming conflicts using the priority order
4. Registration -- Makes skills available to the agent with their declared capabilities and constraints

Skills with triggers in their frontmatter are automatically wired into the scheduler. Skills with requires_tools are checked against the agent's available tools -- if a required tool is not available, the skill is flagged but not blocked.

Agent Self-Authoring

A key differentiator: the agent can write its own skills. When asked to do something it does not know how to do, the agent can use the exec environment to create a SKILL.md and supporting code, then package it as a workspace skill.

Self-Authoring Flow

1. You:   "I need you to check my Notion for new tasks every morning"
2. Agent: Creates skill at ~/.triggerfish/workspace/<agent-id>/skills/notion-tasks/
          Writes SKILL.md with metadata and instructions
          Writes supporting code (notion-tasks.ts)
          Tests the code in the exec environment
3. Agent: Marks the skill as PENDING_APPROVAL
4. You:   Receive notification: "New skill created: notion-tasks. Review and approve?"
5. You:   Approve the skill
6. Agent: Wires the skill into a cron job for daily execution

SECURITY

Agent-authored skills always require owner approval before they become active. The agent cannot self-approve its own skills. This prevents the agent from creating capabilities that bypass your oversight.

Enterprise Controls

In enterprise deployments, additional controls apply to self-authored skills:

• Agent-authored skills always require owner or admin approval
• Skills cannot declare a classification ceiling above the user's clearance
• Network endpoint declarations are audited
• All self-authored skills are logged for compliance review

The Reef

The Reef is Triggerfish's community skill marketplace -- a registry where you can discover, install, publish, and share skills.

Feature	Description
Search and browse	Find skills by category, tag, or popularity
One-command install	triggerfish skill install <name>
Publish	Share your skills with the community
Security scanning	Automated scanning for malicious patterns before listing
Versioning	Skills are versioned with update management
Reviews and ratings	Community feedback on skill quality

CLI Commands

# Search for skills
triggerfish skill search "calendar"

# Install a skill from The Reef
triggerfish skill install google-cal

# List installed skills
triggerfish skill list

# Update all managed skills
triggerfish skill update --all

# Publish a skill to The Reef
triggerfish skill publish

# Remove a skill
triggerfish skill remove google-cal

Security

Skills installed from The Reef go through the same lifecycle as any other integration:

1. Downloaded to the managed skills directory
2. Scanned for malicious patterns (code injection, unauthorized network access, etc.)
3. Enter UNTRUSTED state until you classify them
4. Classified and activated by the owner or admin

INFO

The Reef scans all published skills for known malicious patterns before they are listed. However, you should still review skills before classifying them, especially skills that declare network access or require powerful tools like exec or browser.

Skill Security Summary

• Skills declare their security requirements upfront (classification ceiling, tools, network domains)
• Tool access is gated by policy -- a skill that requires_tools: [browser] will not work if browser access is blocked by policy
• Network domains are enforced -- a skill cannot access endpoints it did not declare
• Agent-authored skills require explicit owner/admin approval
• All skill invocations pass through policy hooks and are fully audited