Research

Web Research

Systematic web research that produces actionable, cited results - not link dumps.

Goal

Produce a concise, sourced answer to the user's question that enables a decision or action. Success = direct answer + cited sources + no unanswered gaps. Failure = link dump, unverified claims, or outdated information.

Dependencies

Tools: WebSearch, WebFetch, Read, Write (all declared in frontmatter) CLI tools: None required Connectors: Public web only — no MCP servers or APIs needed

Context

• No brand voice or codebase conventions needed for general research
• For code-specific research: scan the project's CLAUDE.md or package.json first to understand the tech stack and version constraints before searching
• Reference files in references/ are not required for this skill

Process

Step 1: Understand the Question

Before searching, clarify:

• What specific information do we need?
• What decisions will this inform?
• What would a GOOD answer look like? (docs link? code example? comparison table?)

CHECKPOINT: If the question is ambiguous or the scope is unclear, ask one clarifying question before searching. Do NOT proceed with guesses about intent.

Step 2: Search Strategy

Don't just search once. Use iterative, targeted queries:

Round 1 - Official sources first:

• Search for official documentation: "[tool/library] official docs [topic]"
• Search for GitHub repos: "[tool] site:github.com"
• Search for changelogs/releases if version-specific

Round 2 - Community knowledge:

• Stack Overflow / GitHub issues for specific errors or patterns
• Blog posts from known practitioners (not SEO farms)
• Reddit/HN discussions for real-world experience

Round 3 - Targeted follow-up:

• Fill gaps from rounds 1-2 with specific queries
• Fetch actual pages (WebFetch) for content that search snippets don't cover

Step 3: Source Evaluation

Rank sources by reliability:

Red flags to watch for:

• No dates (could be years outdated)
• No code examples (theory without practice)
• Contradicts official docs
• "Top 10 Best..." listicle format (usually SEO, not substance)
• Content that reads like AI-generated filler

Step 4: Synthesize

Don't dump links. Produce:

1. Direct answer to the question (1-3 paragraphs)
2. Key findings with source attribution
3. Comparison table if evaluating options
4. Code examples from the most reliable source
5. Caveats - what's uncertain, version-dependent, or controversial
6. Sources - links to the actual pages consulted

Step 4 (continued): Present findings

CHECKPOINT: Present the synthesized findings to the user before saving anything. Ask: "Here's what I found. Should I save this to a file, or is this sufficient?"

Step 5: Save (if substantial)

For significant research, write findings to a file:

• RESEARCH.md for general research
• RESEARCH-[topic].md for topic-specific deep dives

This preserves research across context resets.

Output

• Primary deliverable: Inline response with direct answer, key findings, source citations
• Format: Markdown — direct answer paragraph, optional comparison table, code examples if relevant, sources list at bottom
• Save location: RESEARCH.md or RESEARCH-[topic].md in the project root (only if the user confirms or the research is substantial enough to warrant persistence)
• Never save without user confirmation

WebFetch Tips

• Fetch documentation pages directly for accurate information
• Fetch GitHub READMEs for library evaluation
• Fetch specific Stack Overflow answers for verified solutions
• If a page is too large, look for the specific section you need
• Some sites block automated fetching - search results may be your best source

Research Patterns

"What library should I use?" → Search for comparison posts, check GitHub stars/activity, read official docs for each option, produce comparison table with pros/cons/use-cases

"How do I do X in framework Y?" → Official docs first (WebFetch the specific docs page), then community examples, produce code example with explanation

"Why is X happening?" → Search the exact error message, check GitHub issues, read related docs sections, produce root cause + fix

"What's the current best practice for X?" → Search recent (last 12 months) discussions, check if official stance exists, note if community is divided, present consensus with caveats