Skills & Tools
These are two related but different things. Understanding the difference will help you get more out of your assistant.
Tools: the atomic actions
Tools are individual actions your assistant can take. Each one does exactly one thing.
Core tools are always available:
| Tool | What it does |
|---|---|
file_read | Read a file in the workspace |
file_write | Create or overwrite a file in the workspace |
file_edit | Edit a file in the workspace |
bash | Run a shell command (sandboxed to the workspace) |
web_search | Search the internet |
web_fetch | Fetch content from a URL |
memory_manage | Save or update a fact in long-term memory |
memory_recall | Search long-term memory |
skill_load | Activate a skill for the current conversation |
ui_show | Display a card, table, or other visual surface |
Host tools can access your full machine, but require your permission:
| Tool | What it does |
|---|---|
host_file_read | Read any file on your computer |
host_file_write | Write to any file on your computer |
host_file_edit | Edit any file on your computer |
host_bash | Run a shell command on your machine |
Skill tools become available when their skill is loaded. For example, loading the Browser skill adds browser_navigate, browser_click, browser_type, browser_snapshot, browser_extract, and others. Loading the Gmail skill adds tools for archiving, labeling, drafting, sending, unsubscribing, and more.
There are dozens of tools in total. Your assistant picks the right ones based on what you asked for.
Permission levels
Not all tools are created equal. Each tool has a risk level that determines whether your assistant can use it freely or needs to ask first:
- Low risk — read-only operations like searching the web or reading a workspace file. These run automatically.
- Medium risk — stateful changes like writing files or making API calls. These prompt you for approval.
- High risk — destructive or sensitive operations. These require explicit approval with no cached grants.
When a tool needs approval, you'll see an explanation of what it wants to do. Your options:
- Allow — one-time approval
- Allow for 10 minutes — temporary grant
- Always allow — persistent rule (saved to trust.json)
- Don't allow — block this time
- Always deny — persistent block
These rules accumulate over time. The more you use your assistant, the fewer prompts you'll see for actions you've already approved.
Skills: bundles of capability
A skill is a package that combines tools, instructions, and configuration into a coherent capability.
- Tool = "I can navigate to a URL"
- Skill = "I know how to browse the web — navigate pages, click elements, fill forms, extract content, take screenshots, and handle authentication"
Skills give your assistant context about how to use tools for a specific purpose. The Gmail skill knows email workflows. The Browser skill knows how to interact with web pages. The Phone Calls skill knows how to make and receive voice calls through Twilio.
Each skill is a folder containing:
- SKILL.md — instructions that teach the assistant when and how to use the skill
- TOOLS.json — a manifest defining what tools the skill provides, their inputs, risk levels, and execution targets
- tools/ — the implementation code behind each tool
Built-in skills
Your assistant ships with about 30 skills. Here are the main categories:
| Category | Skills | What they enable |
|---|---|---|
| Communication | Gmail, Slack, Messaging, Phone Calls | Read, send, and manage messages across platforms. Make and receive voice calls. |
| Research | Browser, Document, Media Processing, Transcribe | Navigate web pages, write long-form content, process video/audio/images, transcribe recordings. |
| Productivity | Contacts, Tasks, Followups, Notifications, Schedule, Playbooks, Sequences | Track contacts and relationships, manage task queues, set reminders, build trigger-action automations, run email sequences. |
| Computer use | Computer Use, Screen Watch | Control your Mac directly — click, type, navigate between apps. Watch your screen with OCR at intervals. |
| Monitoring | Watcher | Poll external services (Gmail, Calendar, GitHub, Linear) for changes and notify you. |
| Development | Claude Code, App Builder, Orchestration, Subagent | Delegate coding tasks, build interactive web apps, decompose work into parallel subtasks, spawn autonomous background agents. |
| System | Settings, Skill Management, ChatGPT Import | Configure the assistant, create/delete custom skills, import history from other assistants. |
How skills load
Skills aren't all active at once. Your assistant sees a catalog of available skills (names, descriptions, and activation hints) in every conversation. When it determines a skill is relevant — because you asked about email, or said "browse this page" — it calls skill_load to activate it.
Once a skill is loaded:
- Its full instructions (SKILL.md) are injected into the conversation context
- Its tools become available for use
- It stays active for the rest of the conversation
This keeps the assistant's context focused. Only the skills that matter for the current conversation take up space.
Custom skills
If the built-in skills don't cover what you need, you can build your own. Your assistant can help:
"Build me a skill that monitors my favorite subreddit for new posts about TypeScript."
The assistant uses the scaffold_managed_skill tool to generate a full skill package — SKILL.md with instructions, TOOLS.json with tool definitions, and TypeScript executors — then saves it to your skills/ directory. The skill is immediately available for future conversations.
You can also install community skills from external registries, or write skills by hand if you prefer. A skill is just a folder with a few files — there's no compilation step or special tooling required.
The difference, summarized
| Tools | Skills | |
|---|---|---|
| What | Single atomic action | Bundle of tools + instructions |
| Example | browser_navigate | Browser skill (navigate, click, type, extract, screenshot) |
| Scope | Core tools are always available; skill tools load with their skill | Activated per-conversation when relevant |
| Permissions | Gated by risk level (low/medium/high) | Skills themselves don't need permission; their individual tools do |
| Extensible | Via MCP servers or custom skill tools | User-buildable, installable from community |
| Think of it as | A single LEGO brick | A LEGO set with instructions |