Zhihu Public Intel

Overview

Use this skill for Zhihu research and evidence collection. The default keyword/topic/person workflow is check zhihu-mcp runtime -> verify MCP login status -> run visible-browser login helper if needed -> verify again -> use zhihu-mcp tools -> normalize outputs. Do not silently fall back to generic web search or AnySearch before running the zhihu-mcp authentication path unless the user explicitly declines login.

Zhihu often presents login walls for search, article details, comments, and user activity. When that happens, do not stop at "blocked":

1. Record the blocker precisely (signin, 401/403, CAPTCHA, empty tool result, or hidden comments).
2. Guide the user through the local authenticated flow in references/zhihu-mcp-setup.md and scripts/assist_zhihu_login.ps1.
3. Use AnySearch only if the user declines login, two assisted login attempts fail, the MCP runtime remains unavailable, or a public-index cross-check is useful.
4. Never ask the user to paste cookies or request headers in chat.

Mandatory Auth Loop

For keyword/topic/person search, exact-name lookup, comments, user activity, or any task that needs Zhihu search:

1. Run check-runtime.
2. Check login with MCP check_login_status or cookie_status before search.
3. If logged_in=false, login_verified=false, cookies are missing/stale, or a plausible exact-name search returns an auth-looking empty result, immediately run the visible login helper:

powershell -ExecutionPolicy Bypass -File .\skills\zhihu-public-intel\scripts\assist_zhihu_login.ps1

4. Tell the user to complete login/MFA/CAPTCHA in the opened browser, then verify again with check_login_status or cookie_status.
5. Only after verified login, run search_content/detail/comment/profile MCP tools.

Do not merely say "login expired" and switch to AnySearch. Existing cookies.json is not proof of login; if MCP verification fails, treat it as stale and refresh it through the helper. The helper may replace an existing cookies.json only after successful visible login, and it never prints cookie values.

AnySearch is allowed only when:

• the user explicitly declines login or asks for public-index cross-checking;
• the visible helper cannot run in the current environment;
• two assisted login attempts fail or the MCP runtime remains unavailable.

When AnySearch is used, label output as discovery/snippet evidence, not full Zhihu capture.

Decision Tree

1. Identify the target:
   • keyword/topic search;
   • question URL or question ID;
   • answer URL or answer ID;
   • zhuanlan article URL or article ID;
   • comments for an answer/article;
   • user public profile or public recent activity;
   • bulk public crawl.
2. Plan the backend before crawling. For keyword/topic/person targets, the plan must recommend zhihu-mcp-auth:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py plan `
  --target "大模型 Agent" `
  --needs search,question,answers,comments,report `
  --scale medium

3. Execute only the needed lane:
   • zhihu-mcp-auth: default lane. Check runtime, guide visible-browser login if needed, verify login, then use MCP tools.
   • public-browser-lite: detail lane for already discovered public question/answer/article/profile URLs.
   • AnySearch discovery: fallback when user declines login, two assisted login attempts fail, MCP remains unavailable, or public-index cross-checking is requested. Use snippets only as discovery evidence.
   • zhihu-k-search: optional convenience CLI when already installed.
   • zhihu-mcp: authenticated MCP workflow for search, discovered URLs/IDs, full details, comments, or user activity after check-runtime and login-status checks.
   • MediaCrawler: optional larger public bulk crawl.
   • ZhihuApis: optional logged-in comment-completeness lane; cookies stay local.
4. Normalize results before synthesis:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py normalize `
  --input output\zhihu\raw_search.json `
  --source public-browser-lite `
  --output-dir output\zhihu\normalized

5. Produce reports from normalized artifacts, not from untracked browser state.

Backend Routing

Read references/source-routing.md before selecting a backend.

Hard rule for keyword/topic/person tasks:

• Do not run generic web search, direct Zhihu URL probing, or AnySearch before the mandatory auth loop above.
• Run check-runtime; if login is missing/stale or MCP says logged_in=false/login_verified=false, run assist_zhihu_login.ps1 immediately so the user can log in in a visible browser.
• Existing cookies.json can be stale; refresh it through the helper instead of treating the file as success.
• Verify with check_login_status or cookie_status, then use zhihu_mcp.search_content.
• Use AnySearch only after the user declines login, two assisted login attempts fail, MCP remains unavailable, or a public-index cross-check is explicitly requested.

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py check-runtime
powershell -ExecutionPolicy Bypass -File .\skills\zhihu-public-intel\scripts\assist_zhihu_login.ps1

• Use public-browser-lite for already discovered public pages and lightweight reports. Save public raw HTML/JSON-like captures under raw/, then normalize.
• Use zhihu-k-search only as an optional convenience CLI when it is already installed or the user approves setup.
• Use zhihu-mcp as the default after login verification. For repeatable setup, read references/zhihu-mcp-setup.md and prefer scripts/setup_zhihu_mcp.ps1.
• Use AnySearch fallback only after auth is declined/blocked:

python C:\Users\Administrator\.codex\skills\anysearch\scripts\anysearch_cli.py batch_search `
  --query "site:zhihu.com 王皓波" `
  --query "site:zhuanlan.zhihu.com 王皓波" `
  --query '"王皓波" 知乎'

• Use MediaCrawler only when scale matters and the user accepts external crawler setup.
• Use ZhihuApis only when comment completeness is the key requirement and the user approves authorized local cookies.

Common Commands

Print backend plan:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py plan `
  --target "https://www.zhihu.com/question/19550225" `
  --needs question,answers,comments,report

Plan a keyword/person authenticated run:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py plan `
  --target "王皓波" `
  --needs search,articles,comments,report

Extract IDs and classify a URL:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py inspect-url `
  "https://www.zhihu.com/question/19550225/answer/1992353258262504861"

Print the normalized artifact contract:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py schema

Generate a run scaffold:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py scaffold `
  --output-dir output\zhihu\agent_discussion `
  --target "大模型 Agent" `
  --needs search,answers,comments,report

Check private zhihu-mcp runtime without reading cookie values:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py check-runtime

Print the safe login-wall recovery guide:

python skills\zhihu-public-intel\scripts\zhihu_public_intel.py auth-guide `
  --target "王皓波" `
  --needs search,articles,comments

Install/update the optional MCP runtime, then open a visible browser for user-controlled login:

powershell -ExecutionPolicy Bypass -File .\skills\zhihu-public-intel\scripts\setup_zhihu_mcp.ps1
powershell -ExecutionPolicy Bypass -File .\skills\zhihu-public-intel\scripts\assist_zhihu_login.ps1

The login helper stores Playwright-format Zhihu cookies only under the private user-level runtime: %USERPROFILE%\.codex\skills\zhihu-public-intel\runtime\zhihu-mcp\cookies.json. It does not print cookie values.

Output Contract

Keep raw captures and normalized data separate:

• raw/: untouched backend outputs, screenshots, or API captures.
• items.jsonl: normalized question, answer, article, search result, and user rows.
• comments.jsonl: normalized comments and nested replies.
• sources.csv: review table with URL, title, type, author, counts, and backend provenance.
• manifest.json: backend plan, commands run, timestamps, limits, and blockers.
• summary.md: human-facing synthesis with links back to normalized source IDs.

Read references/output-schema.md before merging multiple backend outputs.

Guardrails

• Do not focus on personal favorites, private collections, or browsing history unless the user explicitly asks.
• Do not run generic web search or AnySearch before offering zhihu-mcp auth for keyword/topic/person tasks.
• Default work must remain useful without MCP, paid scraping, external API keys, or committed login state.
• Keep cookies.json, auth.json, browser storage state, .env, request headers, and token-like values out of git and final answers.
• Use logged-in browser state only when the user has authorized access and the target content is available to that account.
• Do not bypass CAPTCHAs automatically or describe blocks as missing data.
• If login is needed, make the user perform login/MFA/CAPTCHA in a visible browser; Codex may run the helper but must not extract or print secrets.
• Rate-limit crawls and prefer small smoke tests before bulk collection.
• Do not republish copyrighted Zhihu content in bulk. Summarize and cite URLs; keep full captures local.
• Respect robots, platform terms, author rights, and legal constraints.

Resources

• scripts/zhihu_public_intel.py: planner, URL inspector, scaffold generator, schema printer, and JSON normalizer.
• scripts/setup_zhihu_mcp.ps1: optional repeatable setup for the zhihu_mcp Codex MCP server in a private global skill runtime.
• scripts/assist_zhihu_login.ps1: optional visible-browser login helper that writes local Playwright-format Zhihu cookies without printing them.
• references/source-routing.md: backend selection and setup notes.
• references/zhihu-mcp-setup.md: details for the optional MCP backend, cookie boundaries, and smoke tests.
• references/output-schema.md: normalized JSONL fields and merge policy.