deep-research

name: deep-research description: Exhaustive, multi-step evidence gathering for systematic reviews, conflicting literatures, and complex multi-hop research questions. license: MIT

Skill: deep-research

Conduct exhaustive, multi-angle research using the model's built-in iterative search. This skill is for retrieval and evidence mapping, not manuscript drafting or autonomous file editing.

When to use

Use this skill when:

• the user asks for deep research, exhaustive coverage, or systematic review
• fast literature review was insufficient
• the evidence base is fragmented, conflicting, or cross-disciplinary
• the question requires multi-hop search and comparison across sources

Do not use this skill when:

• the task is a quick citation lookup
• a small number of sources is enough
• the main task is drafting prose
• the planner only needs a fast relevance check

Operating principles

1. Search broadly first, then narrow.
2. Decompose the research question into subquestions.
3. Track agreement, disagreement, and gaps explicitly.
4. Prefer primary sources, reviews, and open-access sources when possible.
5. Separate source claims from evidence strength.
6. Stop when marginal search yield becomes low.

Search strategy

For each request:

1. Break the problem into 2–6 subquestions.
2. Use MCP tools as primary search (configured globally in opencode.json):

#mcp-academic-search 3. Search each subquestion using MCP tools first. If they fail after reporting the error, use webfetch or direct REST APIs for supplementary coverage. 4. Use multiple angles: core terms, synonyms, competing terminology, adjacent fields, and method-specific queries. 5. Retrieve abstract and DOI for each result via MCP tools. 6. Deduplicate near-identical sources. 7. Prioritize review papers, landmark studies, and directly relevant empirical work. 8. Record uncertainty and unresolved conflicts explicitly.

Stopping criteria

Stop and return results when any of the following is true:

• top sources begin repeating across search angles
• new searches yield mostly duplicates or low-relevance results
• coverage is sufficient to answer all subquestions
• the question cannot be resolved with available evidence

Do not continue searching just because search is available.

Output format

Return a structured research memo.

Header

• Research question
• Scope
• Subquestions searched
• Coverage level: preliminary / moderate / exhaustive
• Known limitations

Per-source note

For each source, provide:

Title: Authors: Year: Venue: DOI/URL: real link or <!-- TODO: verify --> Source type: empirical / review / meta-analysis / theory / preprint Question addressed: Claim: 1-3 sentences describing what the source argues or finds Methods or basis: brief design, dataset, or argument basis Evidence strength: strong / moderate / weak / unclear Relevance: why it matters for the current project Caveats: limitations, ambiguity, or verification issues

Synthesis section

After the per-source notes, include:

• Themes
• Points of agreement
• Points of disagreement
• Missing evidence
• Recommended next search directions

PDF tools

When working with local PDFs:

• pdftotext <file.pdf> - — extract text to stdout for reading
• pdfgrep -i -n "pattern" <file.pdf> — search inside a PDF without converting; useful for locating specific claims, tables, or sections in long papers

Guardrails

#research-guardrails

• Do not write final manuscript prose unless explicitly instructed.
• Do not edit project files unless explicitly instructed elsewhere.
• Mark uncertain metadata with <!-- TODO: verify -->.
• If coverage is weak, say so explicitly.

#mcp-procurement-rule

Handoff rule

Return structured evidence for the planner, writer, or reviewer to use. Do not treat deep research as the final writing stage.

Quick Reference (compact reminders)

• Scope: systematic evidence, multi-hop, conflicting literatures. Not for quick lookups or prose drafting.
• Workflow: decompose → multi-angle search → deduplicate → track agreement/disagreement/gaps → stop when yield collapses.
• Output minimum: research question, subquestions, coverage level, per-source notes, limitations, synthesis.
• Escalation: if only 2–5 papers are needed, recommend finding-refs instead.
• Anti-drift: no fabrication, no manuscript prose unless instructed, stop when marginal yield is low, mark uncertainty with <!-- TODO: verify -->.