Browser Use Launches Browser Skills to Stop AI Agents Relearning 95 Sites
Browser Use launches an open, community-driven directory of site-specific automation playbooks that any coding agent can load with a single prompt
Takeaways
- Browser Skills launched: An open directory of 95 site-specific automation playbooks (domain skills) for AI browser agents, free and open source.
- One-prompt setup: Paste a single prompt into Claude Code, Cursor, Codex, or any supported agent to install and activate all skills instantly.
- Stops redundant discovery: Agents read pre-built markdown playbooks for a site before acting, instead of re-exploring from scratch every run.
- Agent-generated, not hand-written: Skills are created by the harness itself during task execution and contributed back via GitHub PR.
- 95 hosts covered at launch: Including Reddit, Amazon, Gmail, LinkedIn, SEC EDGAR, TikTok, Shopify Admin, arXiv, and more.
- Works everywhere: Compatible with Claude Code, Cursor, Windsurf, Gemini CLI, GitHub Copilot, Cline, and 6 other coding agents. browser-skills.sh
Browser Skills is a new open directory from Browser Use that gives AI coding agents pre-built, site-specific automation playbooks -- called domain skills -- so they stop wasting time figuring out the same websites over and over again. It launched alongside the browser-harness project and is free to use today.
The tax every browser agent pays
Every browser agent does the same thing: it re-discovers every website from scratch, every time it runs. Open a tab, poke around, find the button, click it, parse the response, close the session. Repeat. Agents burn through tokens re-learning sites they have already conquered, and teams painstakingly hand-write scripts for workflows an agent already solved last week -- the same discovery tax paid over and over, across thousands of sessions.
The browser is still the entry point for many real workflows -- admin panels, SaaS dashboards, ecommerce sites, recruiting platforms, CRMs -- and many of them do not expose stable APIs, or their API permissions are harder to obtain than plain webpage access. Browser Skills is the answer to that gap.
What a "domain skill" actually is
A domain skill is a folder of markdown files that captures everything a browser agent needs to know about a specific site: which selectors to use, which flows to follow, which edge cases to watch for, and when to avoid the browser entirely. When BH_DOMAIN_SKILLS=1 is set and a task is site-specific, the agent reads every file in the matching agent-workspace/domain-skills/<site>/ directory before inventing an approach.
The key design decision: skills are written by the harness, not by you. Just run your task with the agent -- when it figures something non-obvious out, it files the skill itself. Hand-authored skill files are discouraged; agent-generated ones reflect what actually works in the browser. The community then contributes those generated files back via pull request.
95 sites, one prompt
The directory currently covers 95 hosts across 105 markdown files, synced directly from the browser-harness repository. The range is broad:
- Scraping: Reddit, YouTube, Amazon, LinkedIn, SEC EDGAR, PubMed, arXiv, Booking.com, Glassdoor, Zillow
- Checkout and booking: AA, Alaska Airlines, Expedia, Ctrip
- Productivity and SaaS: Gmail, Trello, HubSpot, Shopify Admin, Vercel, QuickBooks
- Content and social: TikTok (upload), X (posting), Facebook (groups and pages), Substack
- Finance and data: TradingView, CoinGecko, CoinMarketCap, Polymarket, Macrotrends, World Bank
To activate everything, you paste a single prompt into your coding agent. The prompt tells the agent to install browser-harness if needed, enable domain skills via the BH_DOMAIN_SKILLS=1 flag, and consult the matching skill folder before touching the browser.
Set up https://github.com/browser-use/browser-harness for me if it is not already installed.
If setup is needed, read `install.md` first to install and connect it to my real browser.
Then read `SKILL.md` for normal usage and always read `helpers.py`.
Enable domain skills by setting `BH_DOMAIN_SKILLS=1` for browser-harness.
When my browser task targets a site, use the matching folder under
`agent-workspace/domain-skills/<site>/` and read every markdown file
in that folder before inventing an approach.How browser-harness works under the hood
Browser-harness is described as a thin, editable CDP (Chrome DevTools Protocol) harness for letting LLMs connect to a real browser. CDP is the low-level protocol Chrome exposes for programmatic control -- the same one Playwright and Puppeteer use internally, but here the agent talks to it directly. The agent connects directly to Chrome or Chromium, controls pages through a CDP WebSocket, and can combine screenshots, coordinate clicks, DOM inspection, network requests, and raw CDP commands.
If a helper function is missing in agent-workspace/agent_helpers.py, the agent writes it during execution to solve the task -- and that new helper can later become a domain skill. The harness essentially improves itself as it works. The repo has already accumulated 13.4k stars and 1.2k forks on GitHub.
Works with your existing tools
Browser Skills is not tied to any single agent runtime. It works with Claude Code, Cursor, Windsurf, Cline, Codex, Gemini CLI, GitHub Copilot, Roo Code, Amp, Kiro, Goose, and Trae. Write a skill once, use it everywhere -- this cross-agent compatibility is what makes the ecosystem viable, since you are not betting on a single vendor.
What it is good at -- and where it falls short
Domain skills shine on repetitive, structured tasks against sites with stable layouts: scraping product data, booking flows, filing reports, posting content. A practical path is to start with low-risk, frequent, reversible workflows -- read-only tasks, downloads, organization, and checks -- and once the flow is stable, turn the experience into a skill.
The system has real limits. Domain skills can improve an agent's success rate, but they should not fully automate high-risk operations. Passwords, cookies, tokens, and customer data should never be recorded. Human confirmation should be kept for payments, deletion, bulk submission, account changes, and external publishing. Skills also go stale: they expire after site changes and require revalidation.
The bigger picture
When an agent figures out how to complete a task on a specific website, it generates a skill file that other agents can reuse -- think of it like a social network for agents, where one agent creates a skill and others use it and leave written feedback. That feedback loop is what separates this from a static cheat-sheet library.
The broader trend is clear: AI coding agents are everywhere -- Claude Code, Cursor, GitHub Copilot, Codex, Windsurf, Gemini CLI -- and they all share the same limitation. Out of the box, they know how to write code, but they lack the procedural knowledge -- the step-by-step workflows, conventions, and guardrails -- that turn a generic assistant into a specialist. Browser Skills is a direct attempt to build that shared procedural memory, one site at a time, in the open.
Browser Skills is free and open source. You can browse the full directory at browser-skills.sh, contribute new skills via PR to the browser-harness domain-skills folder, and check the official docs for setup details.