Deep Research

Deep Research

This skill turns a single question into a rigorous, multi-round research process and a structured report with citations. It orchestrates searching, reading, and validation — it does not provide its own search implementation.

When to Use This Skill

• The user asks for in-depth research, a deep dive, or a 深度调研 on a topic
• The question needs current information from multiple sources, not a one-shot answer
• The user wants a report with verifiable citations, not an off-the-cuff summary
• Comparing options, surveying a landscape, or doing due diligence

Do not use this skill for simple fact lookups — a single web search is faster and enough.

Composition with Retrieval Tools

This skill contains no search scripts. Use whatever retrieval tools are available in the current environment, in this order of preference:

1. The web-search skill (if available) for query-based discovery.
2. Built-in web search / web fetch tools exposed by the runtime.
3. The browser tool for login-gated or heavily dynamic pages that plain fetching cannot handle.

Never claim to have searched or read a page unless a retrieval tool was actually invoked. If no retrieval tool is available in the current mode, say so plainly and produce the best report you can from prior knowledge, clearly marked as unverified.

Delegation Protocol (mandatory)

Deep research is not a solo 2–3 round search. When the subagent tool is available, you MUST fan the work out and drive it in loops. Stopping after a couple of quick search rounds is a protocol violation.

1. Clarify the Research Question

Pin down what a good answer looks like:

• Core question: restate it in one sentence. If the request is genuinely ambiguous, ask 1–3 targeted clarifying questions; otherwise proceed with a stated interpretation.
• Scope: time range, geography, industry, depth (brief vs. exhaustive survey).
• Deliverable: report length and required sections.

State your interpretation at the top of the work so the user can correct course early.

2. Plan Angles, Then Fan Out in Parallel

Decompose the question into 3–5 meaningfully different angles (definitions/background, current landscape and key players, recent developments, primary data and statistics, contrarian views and limitations). Then launch one researcher subagent per angle in a single subagent parallel call — do not research the angles one by one yourself.

Each delegation task must be self-contained: the angle, the core question for context, the time range, the languages to search in, and the required output — a list of findings where each finding carries a source URL, publisher, date, and a one-line takeaway.

When the controlled Deep Research shortcut is active, persist the 3–5 angles with workflow_state action plan, and record each source with action source. Those URLs are fetched by the runtime before they count.

3. Cross-Validate the Returns

When the parallel researchers return:

• Merge their source inventories; deduplicate by URL.
• Require at least two independent sources for any load-bearing claim. Mark single-source claims.
• Prefer primary sources over aggregators; check publication dates on fast-moving topics.
• When sources conflict, report the conflict and the credibility of each side — never silently pick one.
• Read the few most load-bearing pages yourself; do not relay snippet-level claims as verified.

4. Loop on the Gaps

Draft the report outline and audit it for gaps: unsupported claims, missing sub-questions, contradictions left unresolved. If material gaps remain, start an agent_loop (goal mode, goal = "all load-bearing claims in the outline are supported by 2+ independent sources") and use each iteration to attack the remaining gaps — with fresh subagent delegations where the gap needs new retrieval. Declare done only when the goal holds or the iterations stop changing the picture. Respect the loop's iteration cap; if it trips, deliver the report with the gaps explicitly listed. In the controlled shortcut, done is only a completion request: it remains active until the recorded angles, researcher delegations, and reachable sources clear the runtime gate.

5. Synthesize the Report

Write the report in the user's language. Recommended structure (adapt to the question):

# [研究主题 / Topic]: Research Report

## 摘要 / Executive Summary
[3–5 sentences: the question, the headline findings, the confidence level]

## 背景 / Background
[Context needed to understand the findings]

## 主要发现 / Key Findings
### Finding 1: [title]
[Evidence and analysis, with inline citations like [1], [2]]

### Finding 2: [title]
...

## 争议与分歧 / Disagreements and Open Questions
[Where sources conflict, what remains unknown]

## 结论 / Conclusions
[Direct answer to the research question; caveats and confidence]

## 来源 / Sources
[1] Publisher — "Title" (date). URL
[2] ...

Citation rules:

• Every non-obvious factual claim carries an inline [n] citation.
• The source list contains full URLs; each source was actually opened/read, not just seen in a results snippet.
• If a section relies on prior knowledge rather than retrieved sources, label it as such.

Quality Bar

Before delivering, check:

• The research was fanned out to parallel researcher subagents (or, if the tool is unavailable, the report says so and explains the degraded process)
• Material gaps were attacked in agent_loop iterations, not abandoned after one round
• The report answers the question actually asked, in the user's language
• Load-bearing claims have 2+ independent sources or are flagged as single-source
• Every citation maps to a real, retrieved URL
• Conflicts and uncertainty are surfaced, not hidden
• No fabricated quotes, numbers, or links
• In the controlled shortcut, save the completed report as a .md file inside the selected workspace and record it with workflow_state role deliverable.
• Save a readable .md, .txt, or .json validation report that audits citations, source conflicts, scope, and remaining uncertainty; record it with role validation. Research evidence alone never completes the shortcut.