nahisaho Deep Research
Iterative hypothesis-driven research engine with Think→Action→Report cycle.
Use This Skill When
- Collecting market, competitive, or technology intelligence.
- Verifying business data from multiple sources.
- Conducting web-based investigation for consulting deliverables.
Research Tools (Priority Order)
Use the following tools in priority order based on availability and task needs:
1. web_search tool (Built-in — Always Available)
The agent's built-in web search tool. Use for targeted queries:
web_search("Japan SaaS market size 2025 CAGR")
web_search("McKinsey digital transformation ROI study")
- Best for: Focused factual queries, market data, statistics, recent news.
- Bilingual: Run each query in both Japanese and English for comprehensive coverage.
2. /research command (CLI — Interactive Sessions)
GitHub Copilot CLI's deep research command for comprehensive investigations:
/research What is the competitive landscape for AI-powered CRM in Japan?
- Best for: Multi-faceted research questions requiring synthesis from many sources.
- Combines: GitHub search + web sources + structured analysis.
- Use when: The research question is complex and benefits from automated sub-question decomposition.
3. deep-research MCP (When Configured)
Structured research via MCP server for full-pipeline research:
- Question Elaboration → refine scope
- Subquestion Generation → decompose into 3-5 sub-questions
- Web Search → collect from authoritative sources
- Content Analysis → evaluate source quality
- Report Generation → structured findings with citations
Workflow
Think Phase
- Analyze knowledge gaps against the true objective.
- Generate hypotheses about likely conclusions.
- Choose tool:
web_searchfor targeted queries,/researchfor comprehensive, MCP for pipeline. - Determine next action: SEARCH / VISIT / VERIFY / COMPLETE.
Action Phase
| Action | Tool | When |
|---|---|---|
| SEARCH | web_search (bilingual JP+EN) |
Need specific data points |
| RESEARCH | /research command |
Need comprehensive multi-source investigation |
| VISIT | web_fetch |
Need full page content from a URL |
| VERIFY | web_search (2nd query) |
Cross-validate key numbers/facts |
| COMPLETE | — | All completion criteria met |
- Bilingual search: Always search in both Japanese and English.
- Academic priority: Prefer
.ac.jp,.edu,.govdomains. - Cross-validation: Verify key data with 2+ sources.
Report Phase
- Update findings after each round.
- Tag each fact with source URL and trust level.
- Mark single-source findings with ⚠️.
Deliverables
results/research-notes.md: findings with source attribution.results/source-registry.md: all sources with trust ratings.
Quality Gates
- 3+ trusted sources collected.
- Weighted trust average ≥ 60%.
- Key hypotheses are verified or refuted.
- Single-source claims marked with ⚠️.
- Results saved to files (not chat-only).
If any gate fails: identify the specific failing check, fix the issue, and re-validate before proceeding.
Gotchas
web_searchは常に利用可能。MCP が設定されていなくても Web 検索は実行できる- 検索は日英両方で行うこと。英語のみだと日本市場固有のデータ（官公庁白書、国内調査）を見落とす
/researchコマンドは CLI インタラクティブセッションでのみ使用可能。自動化パイプラインではweb_searchを使う- ソースの信頼度が「低」のデータ（ブログ、SNS）は単独では使用しない。必ず上位ソースで裏付けを取る
- 収集データは各ラウンド後にファイルに保存すること。コンパクションで中間結果が消失する
- Phase間の引き継ぎは必ず
results/research-notes.mdに保存してからハンドオフすること
Validation Loop
- リサーチラウンドを実行
- チェック:
- 3件以上の信頼できるソースがあるか
- 重要な数値が交差検証されているか
- 単一ソースの主張に ⚠️ が付いているか
- 結果がファイルに保存されているか
- 不合格なら追加ラウンドを実行
- 全基準を満たしてからCOMPLETE