Claude Code gets a browser license

What it does

This is essentially a prompt-and-glue wrapper that turns Claude Code into a browser automation engineer. Packaged as both a Claude Code plugin and a standalone skill, it lets you describe a web task—test a form, screenshot a dashboard, crawl for broken links—and Claude will generate custom Playwright code, execute it through a universal run.js wrapper, and hand back screenshots with console logs. It defaults to a visible, slow-motion browser, so you watch the agent work rather than parse headless logs.

The interesting bit

The project treats context window conservation as a first-class concern: Claude loads only a terse SKILL.md initially, pulling in the full Playwright API reference only when the task demands it. The universal executor exists chiefly to preempt Node module resolution errors, which tells you the author has been burned by require paths before. It also implements the open Agent Skills spec, so it isn’t strictly chained to Anthropic’s ecosystem.

Key highlights

• On-demand code generation: Claude writes task-specific Playwright scripts rather than running pre-baked automations
• Visible-by-default execution (headless: false) with 100ms slow motion for observability
• run.js universal executor handles module resolution to avoid typical Node path errors
• Progressive disclosure: concise skill instructions load first; full API reference is fetched only when needed
• Open Agent Skills spec compliance for portability across agent platforms

Caveats

• The nested plugin structure makes manual installation awkward; you must extract the inner skills/playwright-skill/ folder to the right path
• Troubleshooting section is robust for a reason—missing Playwright installs, module resolution, and browser visibility issues are common enough to document
• It is thin glue code around Playwright plus prompt engineering; the value is in the packaging, not a novel runtime

Verdict

Worth a look if you live in Claude Code and want to outsource quick browser smoke tests or visual checks without writing boilerplate. Skip it if you already maintain a mature Playwright suite or need automation outside the Claude ecosystem.