| name | deep-research |
| description | Research a topic in depth using web search, academic papers, and citation graphs. Use when the user asks to research, investigate, or explore a topic thoroughly. |
| argument-hint | ["topic or question"] |
| allowed-tools | Bash, Read, Glob, Grep, Write |
Research "$ARGUMENTS" in depth using the paper and paper-search CLI tools. Follow this workflow:
1. Broad Discovery
Start with broad searches to map the landscape:
paper-search google web "$ARGUMENTS"
paper-search semanticscholar papers "$ARGUMENTS" --limit 10
Scan titles, snippets, and citation counts. Identify the most relevant papers and key terms.
2. Narrow and Filter
Refine based on what you found:
paper-search semanticscholar papers "<refined query>" --year 2023-2025 --min-citations 10
paper-search semanticscholar snippets "<specific question>"
paper-search pubmed "<query>" # if biomedical
3. Deep Read
For the most relevant papers (at least 3-5), read in depth:
paper outline <arxiv_id> # understand structure first
paper skim <arxiv_id> --lines 3 # quick overview
paper read <arxiv_id> <section> # read key sections
For web sources:
paper-search browse <url>
4. Follow the Citation Graph
For key papers, explore their context:
paper-search semanticscholar citations <paper_id> --limit 10 # who cites this?
paper-search semanticscholar references <paper_id> --limit 10 # what does it build on?
paper-search semanticscholar details <paper_id> # full metadata
5. Synthesize
Combine findings into a structured report with:
- Key findings and themes
- Areas of agreement/disagreement
- Gaps in the literature
- Citations for all claims (include paper titles and URLs)
- BibTeX entries for key papers (use
paper bibtex <arxiv_id> to generate)
Guidelines
- Always start broad, then narrow. Don't read deeply until you've scanned widely.
- Read at least 3-5 primary sources before synthesizing.
- Cross-reference web sources against academic papers when possible.
- Use
paper-search semanticscholar snippets to find specific evidence for claims.
- Track what you've already searched/read to avoid redundancy.
- If a search returns arxiv papers, use
paper read to get the full text rather than just the snippet.
