Research — Source-Quality-Guaranteed Investigation

Academic Papers

Reject: unpublished manuscripts, predatory journal papers, preprints with 0 citations and no institutional backing.

Case Studies & Latest Examples

Reject: anonymous anecdotes, unverified social media claims, press releases without technical detail, secondhand accounts without primary source link.

Phase 1: Scope

Clarify the research scope before searching.

1. Parse the user's request to identify:
   • Topic: what to investigate
   • Type: best practices / papers / case studies / mixed
   • Constraints: time range, specific technologies, industry, scale
2. If the request is ambiguous, ask up to 2 clarifying questions via AskUserQuestion
3. Define 3-5 search queries covering different angles of the topic

Output:

## Research Scope
- Topic: [identified topic]
- Type: [best practices / papers / case studies / mixed]
- Constraints: [time range, tech stack, industry, etc.]
- Search queries: [list of planned queries]

Phase 2: Search & Collect

Execute searches systematically. Breadth first, then depth.

1. Run all planned search queries via WebSearch
2. For each result, assess source reliability tier before reading
3. Prioritize Tier S and A sources — only use Tier B if S/A coverage is insufficient
4. For each promising result, fetch the full page via WebFetch and extract:
   • Key claims and findings
   • Supporting evidence or data
   • Date of publication (check page header, URL path, meta tags, footer — always record this)
   • Author and affiliation
5. Cross-reference claims across multiple sources — a claim backed by 2+ independent sources is stronger

Search Rules

• Execute at least 3 distinct search queries per topic to avoid single-source bias
• If the first page of results is dominated by low-quality sources, reformulate the query (add "site:github.com", "site:arxiv.org", author names, conference names, etc.)
• For papers, prefer searching Google Scholar, Semantic Scholar, or arXiv directly
• For best practices, prefer official documentation sites and established engineering blogs
• Always check publication date — flag anything older than 3 years as potentially outdated (unless it's a foundational work)

Phase 3: Synthesize & Report

Compile findings into a structured markdown report.

# Research Report: [Topic]

## Executive Summary
[2-3 sentence overview of key findings]

## Findings

### [Finding 1 Title]
[Description with specific details, data points, or recommendations]

**Sources:**
- ★★★★★ (理由) [S] Author/Org, "Title", URL, YYYY-MM
- ★★★★☆ (理由) [A] Author/Org, "Title", URL, YYYY-MM

### [Finding 2 Title]
...

## Source Reliability Summary

| Tier | Count | Notes |
|------|-------|-------|
| S    | [N]   | [e.g., "Official docs, top-venue papers"] |
| A    | [N]   | [e.g., "Major tech blogs, peer-reviewed"] |
| B    | [N]   | [e.g., "Community resources — used only where S/A unavailable"] |

## Source List
[Full numbered list of all sources used, with relevance rating, reason, tier, author, title, URL, and date]

## Caveats
- [Any gaps in coverage, areas where only lower-tier sources were available, conflicting findings, etc.]

Report Rules

• Every factual claim must have at least one source citation
• Each source must include its reliability tier in brackets: [S], [A], or [B]
• Each source must include a relevance rating (★1-5) with a short reason explaining how directly the source addresses the research topic:
  • ★★★★★: Directly addresses the core question with specific, actionable detail
  • ★★★★☆: Highly relevant — covers major aspects of the topic
  • ★★★☆☆: Moderately relevant — useful context or partial coverage
  • ★★☆☆☆: Tangentially relevant — related domain but not directly on-topic
  • ★☆☆☆☆: Minimally relevant — only peripherally touches on the topic (avoid including these)
• If findings conflict across sources, present both sides and note the disagreement
• Every source MUST include a publication date (YYYY-MM or YYYY). If the exact date is not visible on the page, check the URL path, meta tags, or page footer for date clues. If truly undeterminable, write "Date unknown" explicitly — never silently omit the date
• The Caveats section is mandatory — even if it's just "No significant caveats"
• Aim for 5-15 sources per report (fewer if the topic is narrow, more if it's broad)

Phase 4: Review

After presenting the report, offer next steps via AskUserQuestion:

1. "Deep-dive into a specific finding" — expand one section with more sources
2. "Search for additional perspectives" — broaden the search to cover gaps
3. "Research complete" — done

Important Notes

• No hallucinated sources: Every URL must come from an actual search result. Never fabricate titles, authors, or URLs
• Recency awareness: Default to preferring recent sources (last 2 years) unless the topic calls for foundational/historical references
• Conflict transparency: When experts disagree, present the disagreement rather than picking a side
• Language: Follow CLAUDE.md preference. Fallback to English if unspecified