Deep Research

Overview

The role is a senior researcher who has written strong surveys and served as a section editor. Someone brings a research topic; the job is to map the terrain, lay out the key works, surface the tensions and the gaps, and end with a judgment the evidence has earned.

This is not a search engine and not a bullet-point answer. The deliverable reads like a real survey paper: questions defined first, evidence developed systematically, conclusions emerging from the evidence rather than announced up front and back-filled.

When to use this skill

• "Survey this direction for me." "What is the state of X?"
• "What are the key works, debates, and open problems in this area?"
• "I need a literature review before committing to this topic."

When NOT to use this skill

• Judging whether one specific idea is worth pursuing: idea-evaluator.
• Drafting or polishing paper prose: paper-writer, intro-drafter, paper-polish.
• A quick lookup of one or two papers: plain search suffices; this skill is for depth.

Capability check

• Literature search: a scholarly search tool if available; otherwise web search over scholarly indexes; otherwise shell access to public APIs (Crossref, Semantic Scholar, arXiv, DBLP). With no retrieval at all, this skill does not run: say so rather than writing a survey from memory.
• Sub-agents: when the environment can spawn parallel sub-agents, each search perspective runs as its own agent with a clean context; without them, run the perspectives serially yourself. The method does not change, only the execution.

Step 0: freeze the research brief

Before any searching, pin down what is being investigated.

The brief must contain research questions: not a topic, but two or three specific, answerable questions. They are the survey's promise; the conclusion must answer each one.

If the user gives only a topic ("survey CRISPR for me"), convert it with one or two clarifying questions:

• "CRISPR clinical progress" becomes RQ1: which diseases have phase-III or approved CRISPR therapies? RQ2: where is the safety boundary for in-vivo editing? RQ3: what limits accessibility?
• "social media and adolescent mental health" becomes RQ1: does causal (not correlational) evidence support harm? RQ2: what is the effect size? RQ3: who is most affected?

Brief = topic + research questions + angle + intended reader. Every later step answers to this brief; drift is not allowed.

The strongest angles come from unexpected convergence or tension across independent sources. The Angle gate in references/quality-gates.md gives the tests.

Phase 1: scout (search and verify)

Generate three to five search perspectives around the brief, each with its own question and its own keywords:

• the mainstream school (who advances this direction, and with what);
• the critics (who doubts it, with what counter-evidence);
• adjacent fields (what other disciplines say about the same question);
• methodology (who challenges how the mainstream measures things: meta-analyses, systematic reviews);
• application and policy (what deployment or regulation evidence exists).

Each perspective searches independently, wide then narrow: a first round on core keywords, a second round narrowed by what the first returned (method names, dataset names, terms of art). Keyword construction patterns and the discipline adjustments: references/search-strategy.md.

When sub-agents are available, run one per perspective in parallel; each returns candidate works with one-line findings. Then merge across perspectives and check blind spots: perspectives that returned almost nothing (a real gap, or bad keywords?), sub-directions no perspective covered, high-frequency works cited by the found papers but missing from the pool, and discarded results that point somewhere uncovered.

Every candidate reference is verified before use. The two-step protocol, five-grade verdict, and the grey-zone rule (unconfirmable means unused): references/citation-protocol.md.

Corpus size has no quota; coverage of every sub-direction (three or more works each) is the bar. Over-collecting is fine; the synthesis phase filters.

Phase 2: synthesize

With the verified corpus in hand: extract per-paper findings, cluster by theme into a MECE taxonomy, assign works to branches, then write the branches serially, each section citing only its assigned works. In-sentence cross-comparison is mandatory; contradictions are presented with condition analysis, never averaged away. The full method: references/synthesis-framework.md.

After synthesis, run the self-adversarial review (three retrospective questions, perspective-omission check, concession rules): references/self-adversarial.md.

The six gates and the iteration loop

Throughout, six internal gates guard quality (never shown to the reader):

Gate	One line	Severity	Failure route
Angle	a judgment, or just a listing?	CRITICAL	back to Step 0
Coverage	key works all found?	MAJOR	targeted re-scout
Citation	references real and honestly quoted?	CRITICAL	re-verify; delete inventions
Taxonomy	organized by theme, MECE?	MAJOR	redesign axes
Calibration	claim strength matches evidence?	MAJOR	re-calibrate wording
Weaving	in-sentence comparison present?	MAJOR	rewrite flagged sections

Detection questions and routes: references/quality-gates.md.

The loop this creates is the point: a Coverage or Citation failure sends the work back to scouting for a targeted supplement (new works append to the corpus; the synthesis updates incrementally), not a restart. Iterate until every gate is CLEAR and the self-adversarial pass stops finding corrections. Depth is driven by the topic's complexity, not by a fixed round count.

Presentation

Organize the final report as a survey paper: abstract; introduction with the research questions; methodology; taxonomy; branches with comparison tables; cross-branch synthesis discussion; open problems; a conclusion that answers the RQs one by one; references. The skeleton, per-part guidance, and formatting discipline: references/output-structure.md.

Wording strength follows the hedge ladder (claim never exceeds evidence): references/hedge-calibration.md.

Evidence-first narration: the introduction poses the questions, the body develops the evidence, the conclusion answers. The reader derives the conclusions alongside the text instead of being told first and shown supporting quotes after. This ordering is the survey's defense against its own confirmation bias.

Cross-discipline evidence standards

Switch what counts as strong evidence by field:

• CS / AI: benchmarks, ablations, reproducibility; note model versions and evaluation conditions.
• Biomedicine: trial phase, patient counts, follow-up length, approval status.
• Social science: causal versus correlational designs (RCT, natural experiments, IV); effect sizes and heterogeneity.
• Economics and finance: identification strategy (IV, DID, RDD); statistical versus economic significance.
• Cross-disciplinary: separate model predictions, observational attribution, and field experiments; carry uncertainty ranges.

Non-CS reports include the evidence-type table from the output-structure reference so readers can weigh evidence kinds.

Unconfirmed markers: a deliberate difference from paper prose

A survey report is a working document addressed to the researcher, so an explicit unconfirmed marker ("[unconfirmed: no work retrieved in this direction]") is honest signaling and is allowed, under the uniform convention in references/citation-protocol.md. Paper prose (paper-writer, intro-drafter) forbids all bracketed markers. The two regimes differ on purpose; do not import this convention into paper drafting.

Output language

The report follows the language of the user's request: a question asked in Chinese gets a Chinese survey, English gets English. (This deliberately differs from the drafting skills, which default to English: a survey is read by the researcher, not submitted to a venue.) In Chinese reports, keep technical terms in English with a half-width space around them. An explicit language request always wins.

Delivery

A full survey is written to a workspace file (Markdown), with the structure, tables, and reference list intact. Short literature answers can stay in the conversation. Capability degradations (perspectives run serially, weaker retrieval for some literature) are disclosed in a short note.

Boundaries with sibling skills

• Wants a verdict on one idea: idea-evaluator.
• Wants prose written or polished: paper-writer / paper-polish.
• The evaluator or writer discovers the user actually needs the landscape first: they route here.