SuperAPI AI Directory
Partner中文AI newsFavorites
Skillssearch-web

search-web

Coding & development

View repository

Updated 10-15 23:23

About this skill

Search-Web Skill

Overview

Integrate with a minimal search chain Context7、Exa、TinyFish, web page body reading,mcp-deepwiki and github-fetcher. For ordinary searches, directly use the tools already available in the current session; when the environment is missing, prompt the user to manually run $search-web setup and restart the code agent.

Principle: use a fixed tool chain for each type of search target; for online searches, prefer Exa, and only use Exa when TinyFish fails, the quota is reached, the service is unavailable, or the result is empty. CLI as a fallback.

Parameter Entry

This skill distinguishes two modes via invocation parameters (i.e., the parameters following /search-web , or the Skill tool's args）：

  • parameter is setup(written as $search-web setup): execute the initialization and MCP configuration process, then stop the current search after completion.
  • Other parameters or no parameters: directly enter the search process, do not run detect.sh, and do not automatically install MCP.

When to Use

  • Query the official documentation of a library, framework, SDK, or language (library docs, API reference, framework guides)
  • Find code examples, tutorials, blog posts, or the latest web materials (code examples, npm packages)
  • Read the body of a specified web page (only after the URL is known)
  • View GitHub repository structure, documentation, and source code entry points (README, source code, directory structure)
  • Read specific files or directory structures in a GitHub repository
  • When encountering library not found、rate limit、documentation outdated and needing to look up the latest information

When Not to Use

  • Purely local code retrieval or ordinary Q&A that does not require internet access
  • Needing to directly execute code, run tests, or debug business logic
  • Complete local documentation already exists and no online verification is needed
  • Needing to modify or refactor code (refactoring)

MCP Quick Reference

Search Target Tool Fixed Rule
Technical Documentation Context7 First resolve-library-id then query-docs, ≤3 times per question
Web Search Exa / TinyFish Exa preferred;Exa use when TinyFish
GitHub Repository Documentation mcp-deepwiki Repository-level documentation overview and in-depth reading
GitHub File/Directory github-fetcher Specific file or directory structure
Web Page Body Host reading tool / TinyFish Use only after the URL is known; when host reading fails, you can use TinyFish to fetch the body

Core Process

  1. If the user parameter is setup, read references/setup.md, execute the initialization process, then stop the current search.
  2. Ordinary searches do not run bash scripts/detect.sh, and directly select tools according to the search target.
  3. For configuration-type questions (threshold, switch, flag, option, parameter, env), first read references/strategies/config-index-shortcut.md。
  4. Technical documentation first goes through Context7(first resolve-library-id then query-docs）。
  5. Online searches first use Exa; when Exa When it fails, hits a quota limit, is unavailable, or returns empty results, follow references/tinyfish-fallback.md use TinyFish CLI as a fallback.
  6. When search hits a GitHub repository, use mcp-deepwiki; when a specific file or directory needs to be read, use github-fetcher。
  7. Only when the user has already given an explicit URL,Exa a URL has been located, or TinyFish a URL has been located, is reading the webpage body permitted.
  8. Finally, organize the answer according to references/rules/output-format.md .

MCP usage boundaries

For the tool mapping, see the "MCP Quick Reference" table above. Fixed prohibitions:

  • Do not use open-websearch or any other unspecified search tool in place of Exa / TinyFish
  • Do not use mcp-deepwiki in place of github-fetcher file reading, or vice versa
  • In normal search, ifContext7、Exa、mcp-deepwiki、github-fetcher reports not registered, not loaded, connection failure, or server unavailable , stop the current search, and prompt the user to run $search-web setup and restart the code agent.
  • Normal search does not automatically execute setup-mcp.sh, and does not automatically write any MCP configuration.
  • TinyFish is a CLI fallback tool; only try to use it when Exa has been called but failed, hit a quota limit, is unavailable, or returned empty results.
  • When the host type is unrecognized, normal search proceeds as usual; the user may proactively run $search-web setup to obtain the generic MCP JSON.

Common errors

  • Skipping resolve-library-id and directly guessing libraryId(Context7 must be called in two steps)
  • Calling libraryName while missing the resolve-library-id
  • Passing libraryName to query-docs(should use libraryId）
  • Exa When the MCP is not registered or not loaded, directly use TinyFish(should be handled as an MCP missing and prompt a restart; only when Exa search calls return failure, hit a quota limit, are unavailable, or return empty results should TinyFish）
  • Treating webpage body reading as a search step in place of Exa
  • Using mcp-deepwiki in place of github-fetcher to read files, or vice versa
  • Continuing to call Context7 after exceeding the 3-call limit (should switch to Exa)
  • Repeatedly retrying when the MCP is not loaded (should prompt the user to run $search-web setup and restart)

Continue loading on demand

Only read examples/usage.md：

  • when you need Context7、Exa、TinyFish, webpage body reading,mcp-deepwiki or github-fetcher parameter examples
  • When you need Context7 error demonstrations, call-order examples, or multi-source integration cases
  • When you need webpage body reading fallback cases
  • When you need common failure scenarios or troubleshooting examples

If the example documents conflict with the main process, this file prevails.

Only when the user calls $search-web setup, or the user explicitly requests configuring/fixing the search environment, read references/setup.md。

When you need standard output templates, source annotation rules, or Markdown structure constraints, read references/rules/output-format.md。

When you need TinyFish installation,TINYFISH_API_KEY writing, fallback search, or webpage body fetching commands, read references/tinyfish-fallback.md。

When encountering questions about exact configuration item names, thresholds, switches,flag、option、parameter、env and similar, read references/strategies/config-index-shortcut.md。

Output requirements

  • Before the final answer, you must format according to references/rules/output-format.md format
  • Clearly annotate the source type:Context7、Exa、TinyFish, webpage body reading,mcp-deepwiki、github-fetcher
  • Preserve original links, document entry points, or repository identifiers
  • If downgrading from Exa to TinyFish, you must state Exa 's reason for failure
  • If a downgrade, read failure, or insufficient information occurs, state it directly
  • Do not fabricate content that was not read