Skip to main content
Skills let you teach the BrowserOS agent how to handle specific tasks. Each skill is a set of instructions written in plain Markdown that the agent loads when it recognizes a matching task. Think of skills as recipes: you write the steps once, and the agent follows them whenever that type of task comes up. BrowserOS implements the open Agent Skills specification, so skills you create are portable across any AI agent that supports the standard.

​
How Skills Work

1

You create a skill

Give it a name, a short description of when to use it, and write the instructions in Markdown.
2

The agent sees the skill catalog

When a conversation starts, the agent loads a list of all your enabled skills with their names and descriptions.
3

The agent matches a task

When your request matches a skill’s description, the agent loads that skill’s full instructions and follows them.

​
Creating a Skill

1

Open Skills settings

Go to Settings and click Skills in the sidebar.
2

Click New Skill

Click the New Skill button to open the creation form.
3

Fill in the details

  • Name: A short, descriptive name (e.g., “Morning Status Report”)
  • Description: Tell the agent when to use this skill. Be specific. For example: “When the user wants to read status updates from work across Notion, Linear, and Slack”
  • Content: Write your instructions in Markdown. Include step-by-step directions, examples, and edge cases.
4

Save and enable

Click Create. The skill is enabled by default and will be available to the agent immediately.
Write your description like a trigger. The agent uses it to decide whether to activate the skill. A good description says both what the skill does and when to use it.

​
Example Skills

Description: When the user wants to read status updates from workInstructions:
Always look for updates in 3 sources:
1. **Notion** - Check the team updates page for any new entries from today
2. **Linear** - Look at issues assigned to the user that were updated in the last 24 hours
3. **Slack** - Check the #team-updates and #engineering channels for unread messages

Summarize everything in a single report grouped by source.
If a source has no updates, say so.
Description: Extract text and tables from PDF files, fill PDF forms, and merge multiple PDFs. Use when the user mentions PDFs, forms, or document extraction.Instructions:
When extracting text from a PDF:
1. Download or open the PDF in the browser
2. Use the page content tool to extract visible text
3. Preserve table structure using Markdown tables
4. If the PDF has multiple pages, process each page

When filling a PDF form:
- Ask the user for the values if not provided
- Fill each field carefully and confirm before submitting

See references/FORMS.md for common form templates.
Description: When the user asks to review code, a pull request, or wants feedback on code qualityInstructions:
Follow this checklist for every code review:
1. Check for security issues (XSS, injection, hardcoded secrets)
2. Look for performance problems (N+1 queries, unnecessary re-renders)
3. Verify error handling is present and meaningful
4. Check that naming is clear and consistent
5. Look for missing tests for new logic

Format your review as a list of findings with severity: Critical, Warning, or Suggestion.
Always start with what the code does well.

​
Managing Skills

From the Skills settings page, you can:
  • Enable or disable a skill using the toggle switch. Disabled skills are not loaded by the agent.
  • Edit a skill’s name, description, or instructions by clicking the edit icon.
  • Delete a skill by clicking the trash icon. This removes the skill permanently.

​
Skill File Format

Under the hood, each skill is stored as a SKILL.md file following the Agent Skills specification: 
---
name: morning-status-report
description: When the user wants to read status updates from work
metadata:
  display-name: Morning Status Report
  enabled: "true"
---
Always look for updates in 3 sources:
1. Notion - Check the team updates page
2. Linear - Look at assigned issues updated in the last 24 hours
3. Slack - Check #team-updates and #engineering channels

Summarize everything in a single report grouped by source.
The file uses YAML frontmatter for metadata and Markdown for the instructions.

​
Frontmatter fields

FieldRequiredDescription
nameYesLowercase, hyphenated identifier (e.g., morning-status-report)
descriptionYesWhen and how the agent should use this skill
licenseNoLicense for the skill
compatibilityNoEnvironment requirements
metadataNoExtra fields like display-name, enabled, version
allowed-toolsNoRestrict which tools the skill can use (experimental)

​
Supporting files

A skill can include additional directories alongside SKILL.md:
  • scripts/ for executable code the agent can run
  • references/ for detailed documentation loaded on demand
  • assets/ for templates, images, or data files
morning-status-report/
├── SKILL.md
├── scripts/
│   └── format-report.py
└── references/
    └── REFERENCE.md
The agent loads the main SKILL.md first. Supporting files are only loaded when the instructions reference them, keeping context usage efficient.

​
Where Skills Live

Skills are stored as folders inside your BrowserOS configuration directory:
OSPath
macOS~/.browseros/skills/
Windows%USERPROFILE%\.browseros\skills\
Linux~/.browseros/skills/
Each skill gets its own folder named after the skill’s name field.

​
Tips for Writing Good Skills

Be specific in descriptions

Include keywords the agent can match against. “When the user asks about PDFs, forms, or document extraction” is better than “Helps with documents.”

Keep instructions focused

A skill should do one thing well. Split complex workflows into multiple skills rather than one large one.

Include examples

Show the agent what good output looks like. Examples reduce ambiguity and improve results.

Use supporting files

Move detailed references to separate files. The agent loads them only when needed, saving context space.
Skills follow the open Agent Skills specification. Skills you create in BrowserOS work with any agent that supports the standard.