Guides
Research workflows
How an agent gets from a first question to a saved, annotated reading list with Scholar.
# Research workflows The agent picks the tools itself. Whether the result holds up depends on a few habits: search widely, screen quickly, read closely, and keep what survives. ## Search with the question itself `search_papers` is built for whole questions. Give it the question as you would ask a colleague, a sentence you remember, or a title. Semantic sources receive the sentence and keyword sources receive keywords drawn from it, so rewriting it into keywords first usually loses information. Narrow the search with parameters, not with wording. - `year="2024-2026"` or `year="2023"` sets the years, and `venue="NeurIPS"` the venue. - `min_citations`, `paper_types` and `open_access_only` make the shortlist stricter. - `intent="survey"`, `"foundational"`, `"recent"`, `"method"` or `"dataset"` leans the ranking toward one kind of need while relevance still counts. - `sort="citations"` or `sort="date"` replaces the ranking outright, so use it only when you want a hard order. Ask in the language of the literature you want. For computer science and AI that is usually English. ## Screen, then read Each result has a title, authors, venue, citation count, a short abstract and an `id` for follow-up calls. Any identifier Scholar returns, a paper link or an exact title works in every tool. Screen with this first, then open the papers that matter with `read_paper`. - The default `pages="1-10"` usually reaches the conclusion of an AI conference paper. - Use `pages="11-20"` for references and appendices, up to 20 pages per call. A long range stops at a page boundary after 100,000 characters, and `next_pages` gives the next range. - Every read lists the figures and tables by name. Read again with one of them, such as `visual="Figure 3"`, and the agent sees that figure with the text around it. `read_paper` can reuse a retained PDF; a new temporary copy is removed after reading. Use `download_paper` to keep one. ## Follow citations `paper_info` returns detail, citations and references in one call with `include="detail,citations,references"`. Citations are later papers that cite it, references are the papers it cites, and both lists are ordered by citation count. `recommend_papers` looks sideways. `relation="similar"` finds papers with similar content, `"peers"` finds papers often cited together with it, and `"kin"` finds papers that cite the same works, which often crosses fields. ## Keep what matters `download_paper` stores a checked PDF in your papers folder and records it in a library collection (`downloads` unless you name another). With the research profile, `paper_library` saves papers, keeps notes and tags, searches what you saved and exports a collection to Obsidian. Saved work lasts across sessions, so the next conversation can start from the library. See [Paper library](/scholar-mcp/paper-library/). ## Map a field `build_paper_graph` starts from one or a few seed papers, follows citations outward and returns nodes, edges, PageRank, bridge papers and a Mermaid diagram. Start with one hop and about 30 papers. When the shape looks right, raise `max_hops` (up to 3) or `max_papers` (up to 100) and give a `topic_filter` so a large graph stays on topic. <Aside type="tip" title="Deep Research skill"> The Claude Code and Codex plugins include a Deep Research skill. It chains searching, reading papers, walking the citation graph and saving chosen papers to the library into one loop that keeps a map of the field up to date. </Aside> ## When results look thin If a source did not answer and that affected the results, the response carries a short warning. Add `debug=true` to see each source's status, count and time and the ranking model that ran, or read `scholar://status`. The [FAQ](/scholar-mcp/faq/) covers slow searches and missing sources.The agent picks the tools itself. Whether the result holds up depends on a few habits: search widely, screen quickly, read closely, and keep what survives.
Search with the question itself
Section titled “Search with the question itself”search_papers is built for whole questions. Give it the question as you would ask a colleague, a sentence you remember, or a title. Semantic sources receive the sentence and keyword sources receive keywords drawn from it, so rewriting it into keywords first usually loses information.
Narrow the search with parameters, not with wording.
year="2024-2026"oryear="2023"sets the years, andvenue="NeurIPS"the venue.min_citations,paper_typesandopen_access_onlymake the shortlist stricter.intent="survey","foundational","recent","method"or"dataset"leans the ranking toward one kind of need while relevance still counts.sort="citations"orsort="date"replaces the ranking outright, so use it only when you want a hard order.
Ask in the language of the literature you want. For computer science and AI that is usually English.
Screen, then read
Section titled “Screen, then read”Each result has a title, authors, venue, citation count, a short abstract and an id for follow-up calls. Any identifier Scholar returns, a paper link or an exact title works in every tool. Screen with this first, then open the papers that matter with read_paper.
- The default
pages="1-10"usually reaches the conclusion of an AI conference paper. - Use
pages="11-20"for references and appendices, up to 20 pages per call. A long range stops at a page boundary after 100,000 characters, andnext_pagesgives the next range. - Every read lists the figures and tables by name. Read again with one of them, such as
visual="Figure 3", and the agent sees that figure with the text around it.
read_paper can reuse a retained PDF; a new temporary copy is removed after reading. Use download_paper to keep one.
Follow citations
Section titled “Follow citations”paper_info returns detail, citations and references in one call with include="detail,citations,references". Citations are later papers that cite it, references are the papers it cites, and both lists are ordered by citation count.
recommend_papers looks sideways. relation="similar" finds papers with similar content, "peers" finds papers often cited together with it, and "kin" finds papers that cite the same works, which often crosses fields.
Keep what matters
Section titled “Keep what matters”download_paper stores a checked PDF in your papers folder and records it in a library collection (downloads unless you name another). With the research profile, paper_library saves papers, keeps notes and tags, searches what you saved and exports a collection to Obsidian. Saved work lasts across sessions, so the next conversation can start from the library. See Paper library.
Map a field
Section titled “Map a field”build_paper_graph starts from one or a few seed papers, follows citations outward and returns nodes, edges, PageRank, bridge papers and a Mermaid diagram. Start with one hop and about 30 papers. When the shape looks right, raise max_hops (up to 3) or max_papers (up to 100) and give a topic_filter so a large graph stays on topic.
When results look thin
Section titled “When results look thin”If a source did not answer and that affected the results, the response carries a short warning. Add debug=true to see each source’s status, count and time and the ranking model that ran, or read scholar://status. The FAQ covers slow searches and missing sources.