一键导入
playwright-mcp-browser-search-optimization
Optimization guide for extracting information and using MCP browser tools effectively.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Optimization guide for extracting information and using MCP browser tools effectively.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Label video behavior segments from frame-grid images with a defined behavior list, model-ready JSON, resumable outputs, and no_behavior handling for sparse labels.
Use Annolid GUI tools for robust automation across web, PDF, video, and chat controls.
Infer the assay or paradigm from video context, task text, tracked entities, and experimental cues such as social interaction, open field, courtship, resident-intruder, and novel object recognition.
Choose assay-specific features and measurable objectives for behavior analysis, including distances, zones, speed, contact, orientation, and object interaction.
Segment behavior timelines from tracks, pose, contact, speed, and proximity signals into typed intervals with stable labels and rationales.
Keep behavior analysis aligned with assay objectives, controls, reproducibility, and measurable outputs instead of ad hoc summaries.
| name | Playwright MCP & Browser Search Optimization |
| description | Optimization guide for extracting information and using MCP browser tools effectively. |
[!WARNING] MCP Tools are Secondary Backups for Web Browsing Always prioritize native Annolid tools (
web_search,web_fetch,gui_web_run_steps) for web interaction before falling back to Playwright or MCP tools. MCP tools should only be used if native tools fail or cannot handle highly complex JavaScript evaluations.
When you need to extract information from websites using MCP browser tools or headless evaluation scripts (like mcp_github_com_microsoft_playwright_mcp_browser_evaluate), it is critical to avoid returning full raw HTML or massive JSON blobs back to your context.
When using evaluation tools, do NOT return document.documentElement.outerHTML or massive node objects. The Annolid MCP tool wrapper will forcefully truncate dicts/lists > 50 items and strings > 50,000 characters to prevent context overflow.
Always prefer returning clean text:
document.body.innerTextdocument.querySelector('main').innerTextIf you need to extract many items (e.g. search results, tables), map them to a simple Array of objects or strings within the Javascript context before returning.
DO NOT DO THIS:
// BAD: Returns massive DOM Node references or breaks serialization
document.querySelectorAll('.product-item')
DO THIS:
// GOOD: Returns a clean list of strings or slim JSON objects
Array.from(document.querySelectorAll('.product-item')).map(el => {
return {
title: el.querySelector('h2')?.innerText || '',
price: el.querySelector('.price')?.innerText || '',
link: el.querySelector('a')?.href || ''
};
})
If you are using the Playwright MCP server's click or fill tools, you must use standard CSS selectors or Playwright text selectors.
"button:has-text('Submit')"'#login-form input[name="username"]''.shopping-cart'If your command returns large data, annolid will intercept it:
"... (X more items truncated)"[WARNING: MCP tool response was truncated...] message appended.If you see these warnings, do not attempt to fetch the same data again unless you change your script to target a more specific selector or paginate the results via Javascript (e.g. .slice(50, 100)).