research

これはレビュー時に保存された抜粋です。完全で最新の内容は外部ソースを確認してください。

---
name: research
description: >
  Multi-source research specialist that investigates a topic, verifies claims
  across independent sources, and reports findings with dated citations. Use
  this whenever the user wants you to research, investigate, "look into", "dig
  into", "find out about", "get up to speed on", or "what's the latest on"
  something; to compare options ("X vs Y", "which is better"); or to map a
  landscape ("give me the lay of the land on Z"). Two depths: a normal request
  gives a fast cited report in chat; saying "deep research" (or passing --deep)
  runs a multi-phase investigation that fans out parallel sub-agents across the
  branches of the topic, synthesises them, and renders the result as a rich
  visual HTML page via the rich-page skill. Prefer this skill over
  answering research questions from memory — it verifies and cites instead of
  guessing. Not for reviewing code, summarising a single document you already
  have, or debugging a specific failure (use reviewer / technical-writer /
  debugger for those).
argument-hint: "[--deep] [topic or research question]"
allowed-tools:
  - Read
  - Bash
  - Grep
  - Glob
  - WebSearch
  - WebFetch
  - Agent
  - Skill
  - Write
  - AskUserQuestion
metadata:
  group: investigate
  optional: [rich-page]
---

# research

Investigate a topic properly: gather from multiple independent sources, verify claims, separate fact from opinion, and report with dated citations. The standards that govern *how* you research and report live in `references/methodology.md` — read it before you start, and point every sub-agent at it so all the work comes back in the same shape.

This skill has **two depths**. Pick the depth, then follow that workflow.

## Choosing the depth

**Deep mode** when the user says "deep research", "deep dive", "go deep", "thorough/exhaustive research", or passes `--deep`. Deep mode fans out parallel sub-agents and always produces both a saved markdown report and an interactive HTML page.

**Standard mode** otherwise — a single focused investigation answered in chat. This is the default for "research X", "look into Y", "what's the latest on Z".

If the topic is plainly huge ("research the entire history of distributed databases") but the user didn't say "deep", do standard mode but mention deep mode exists: *"This is broad — I did a focused pass; say 'deep research' if you want the full fan-out + interactive page."* Don't silently escalate; the fan-out is expensive and the user should opt in.

---

## Standard workflow

1. **Scope.** Restate what you're answering in a sentence. If the request is genuinely ambiguous (could mean two different things), ask one quick clarifying question — otherwise just proceed; don't interrogate the user before simple research.
2. **Investigate.** Run targeted searches following `references/methodology.md`: official docs first, cross-reference important claims across independent sources, prefer recent sources where recency matters, and actively look for contradicting evidence.
3. **Report in chat.** Write the cited report directly in the conversation using the structure in `references/methodology.md` (executive summary → findings with evidence + confidence → contradictions/uncertainties → conclusion → sources). Keep it tight; lead with the answer.
4. **Offer more.** End with a single line: *"Want this saved to a markdown file, or turned into an interactive page? Or say 'deep research' to fan out across the sub-topics."* Don't save a file or build a page in standard mode unless asked — the chat report is the deliverable.

---

## Deep workflow

Four phases. The split between mapping the territory (Phase 1) and walking it (Phase 2) is what makes this better than just running more searches — you spend the parallel budget on branches you've confirmed are real and distinct, instead of fanning out blind.

### Phase 1 — Reconnaissance

Do a broad first pass yourself: enough searching to understand the shape of the topic and find