web-research

name	web-research
description	Search the internet to find solutions, documentation, tutorials, and best practices when you don't know how to do something. Use this when facing unfamiliar APIs, libraries, errors, or techniques that require up-to-date knowledge beyond the training cutoff.
disable-model-invocation	false
allowed-tools	WebSearch, WebFetch, Read, Write, Edit, Bash

Web Research — Internet Search for Solutions

Search the web when you encounter something you don't know how to do, an unfamiliar error, or need current documentation.

When to Use

• Unfamiliar library, API, or framework
• Error message you can't resolve from context alone
• Need current docs (versions, breaking changes, deprecations)
• Looking for best practices or architectural patterns
• Comparing tools, libraries, or approaches

Critical Rules

1. Search first, guess never — if you're unsure, search before writing code
2. Verify sources — prefer official docs, GitHub repos, and Stack Overflow over random blogs
3. Extract only what's needed — don't dump entire pages, summarize the relevant solution
4. Cite sources — always tell the user where the info came from
5. Cross-reference — for critical decisions, check at least 2 sources

Phase 1: Search

Use WebSearch with a targeted query. Tips for good queries:

• Include the library name + version when relevant
• Add language (python, typescript, etc.)
• Include the error message (quoted) for debugging
• Add the current year for recent results

Examples

Scenario	Query
Unknown API	"FastAPI WebSocket" authentication middleware 2026
Error	python "ModuleNotFoundError" anthropic pip install
How-to	ollama openai-compatible API streaming python example
Architecture	multi-agent orchestration patterns LangGraph vs custom
Comparison	vLLM vs Ollama performance M1 Mac 2026

Phase 2: Deep Dive (if needed)

If search results point to a useful page, use WebFetch to read it:

WebFetch(url="https://docs.example.com/guide", prompt="Extract the setup instructions and code examples for X")

Use focused prompts with WebFetch — ask for exactly what you need, not "summarize everything".

Phase 3: Apply

1. Summarize the solution in 2-3 sentences for the user
2. Show the source — include the URL
3. Adapt the solution to the project's conventions (English, existing patterns, file structure)
4. Implement if the user asked for a fix, or propose if it's a design decision

Fetching Twitter/X Content

X/Twitter blocks direct scraping. Use the vxtwitter API to fetch tweet content:

Step 1: Convert the URL

Replace x.com or twitter.com with api.vxtwitter.com:

Original:  https://x.com/username/status/1234567890
API URL:   https://api.vxtwitter.com/username/status/1234567890

Step 2: Fetch via WebFetch

WebFetch(
  url="https://api.vxtwitter.com/username/status/1234567890",
  prompt="Return the complete raw content: tweet text, username, date, any URLs or media links."
)

The API returns JSON with: text, user_name, date, media (image/video URLs), likes, retweets, replies.

Step 3: Read images (if present)

If the tweet contains images with text (infographics, lists, code screenshots), use the Read tool on the image URLs from the media array to extract their content.

Examples

Want	Do
Read a tweet	WebFetch on api.vxtwitter.com/user/status/ID
Read a thread	Fetch each tweet in the thread by ID
Search tweets	WebSearch with site:x.com "keyword" or from:username keyword
Read tweet images	Get media URLs from API response, then Read them

Fallback chain

If api.vxtwitter.com fails:

1. Try WebFetch on https://vxtwitter.com/user/status/ID (HTML version)
2. Try WebSearch with the tweet URL or from:username + keywords
3. Try WebFetch on https://threadreaderapp.com/user/username for threads

Common Research Patterns

Debugging an Error

1. Search: "exact error message" language framework
2. Check top 3 results for matching context
3. If Stack Overflow: read the accepted answer + comments
4. If GitHub issue: check if there's a fix or workaround

Learning a New API

1. Search: library-name getting started guide
2. WebFetch the official quickstart page
3. Search: library-name + specific-feature example
4. Adapt examples to project code

Choosing Between Options

1. Search: option-A vs option-B use-case year
2. Look for benchmark comparisons
3. Check GitHub stars, last update, maintenance status
4. Recommend with trade-offs explained

Output Format

When reporting findings to the user:

**Found**: [brief description of solution]

[Implementation or code snippet]

Source: [URL]

If nothing useful found after 3 searches, say so honestly and suggest alternative approaches.