autoverse-seed-expand

Autoverse Seed Expand

Help students start from a given paper or a given author to sort out the references, subsequent citations, similar work of that research, or other works by that author, and write a manuscript and an evidence table. The Agent is responsible for raising questions, formulating search queries, comparing the literature and writing the main text. Autoverse is responsible for retrieval, paper identification and verification of bibliographic records. The Agent does not write separate retrieval programs.

Procedure

• Before starting retrieval, the Agent first runs autoverse --json --quiet whoami. Only when this command returns a logged-in account or platform key may the Agent continue. When this command returns NOT_AUTHENTICATED or shows that the user is not logged in, the Agent stops retrieval and tells the user to run autoverse login or set AUTOVERSE_API_KEY, and continues only after the user confirms. The Agent must not open the login flow on the user's behalf.
• The Agent first confirms that what the user wants is to expand citations along a given paper, or to find works along a given author, rather than writing a literature review, and not to organize research understanding before a thesis proposal. For scope not specified by the user, the Agent continues according to the following defaults and states the defaults used in the conversation. The Agent need not ask another round of questions for information already provided.
• The defaults are as follows. The Agent focuses on literature from the past five years, while still including the foundational research on the topic. The Agent makes judgments based only on abstracts. The Agent submits a manuscript (.md) and an evidence table (.csv), and does not submit PDFs.
• When the user provides a DOI, PMID, PMCID, arXiv number or Semantic Scholar ID, the Agent uses resolve to identify the paper. When the user provides only a title, a short name, or when the identity of the item is still uncertain, the Agent first searches with search (combined with --author or --year-from for precise disambiguation when necessary), and identifies the paper after comparing title, authors and year. When the user provides an author name, the Agent uses authors to identify the person, and continues after comparing the returned author records. When the user provides an ORCID, the Agent still calls authorswith the name, and verifies the author's identity with that ORCID in the conversation. The Agent must not fabricate other flags of authors . When the user first provides a paper and then asks to see other works by that paper's author, the Agent first identifies the paper, obtains the author name from resolve or batch , then uses authors to identify the person, and then searches for other works by that author.
• When the starting object is a paper, the Agent searches references, subsequent citations or similar work separately according to the questions the user cares about. When the starting object is an author, the Agent first confirms the author's identity, then searches for that author's works, and for valuable items continues to search references, subsequent citations or similar work. When authors with the same name have not yet been clearly distinguished, the Agent first checks that author's affiliation, collaborators or representative works. The Agent must not write different people's works into the same list. The Agent revises search queries based on retrieval results. The Agent must not substitute the number of calls or retrieval rankings for judgment about the literature.
• For parameters other than the following examples, or when a command cannot be executed, the Agent then reads autoverse-cli or the command description. The Agent does not guess option names.
• For discovering literature, identifying papers and identifying people, the Agent uses only the following Autoverse commands. The Agent must not use web search or other sites as an entry point for discovering literature. The Agent must not open publisher pages to supplement bibliographic records.
• When an Agent includes two or more papers, write the identifiers into ids.txt, and use batch to identify the papers in one go. The Agent must not identify each paper separately with resolve。

Common calls:

autoverse --json --quiet whoami
autoverse --json --quiet search "<title or topic>"
autoverse --json --quiet search "<title>" --author "<author name>"
autoverse --json --quiet authors "<author name>"
autoverse --json --quiet resolve <prefixed-id>
autoverse --json --quiet batch -f ids.txt
autoverse --json --quiet related <seed-id> --via references
autoverse --json --quiet related <seed-id> --via cited-by
autoverse --json --quiet related <seed-id> --via similar

When the curated command does not return citation counts or bibliographic details, the Agent may use only the following paths, invoked via api . When they duplicate a curated command, the Agent still uses the curated command. The Agent must not call paths not listed here. The Agent must not make direct HTTP requests. The Agent must not consult the OpenAPI on its own.

autoverse --json --quiet api -X GET /v1/papers/search
autoverse --json --quiet api -X GET /v1/papers/resolve
autoverse --json --quiet api -X POST /v1/papers/batch
autoverse --json --quiet api -X GET /v1/semantic-scholar/native/graph/paper/{paper_id}
autoverse --json --quiet api -X GET /v1/usage

Citations and Evidence

Claims and literature points in the manuscript must be traceable to entries verified by Autoverse. An abstract supports only the judgments the abstract can support. Identifiers written into the manuscript or the evidence table must be identified via resolve or batch . Papers that cannot be identified must not be treated by the Agent as cited literature. The Agent must not fabricate PMIDs, DOIs, titles, authors, or research conclusions.

Short abstracts in bibliographic records are for screening only. Before writing, the Agent must read the full abstract and the fields already returned. Authors, year, conference or journal are taken from search、resolve or batch. Indexing catalogs, rating facts, and OA links are returned in full in resolve and batch .search If only the first author is given, the Agent uses resolve or batch to obtain all authors. Citation counts and document type: if the curated command already provides them, use them; if not, use the above api. Leave still-missing fields blank. The Agent must not write "not returned". The Agent must not open web pages to fill in gaps. The Agent must not sort by citation count.

When there is no lawful open-access full text, judgment stops at the abstract. The Agent must not bypass subscriptions or paywalls.

Citation relationships already returned by Autoverse may be written by the Agent as reference relationships or subsequent citation relationships between two specific papers, marked respectively with references and cited-by . Related work, research continuation, or subsequent works by the same author that the Agent judges based on abstract content must be stated as content-level relatedness and must not be written as verified citations. When evidence is insufficient or literature conclusions are inconsistent, the Agent must state this in the discussion and must not smooth it over. When retrieval is temporarily unavailable, the Agent must explain the coverage scope in the literature sources and must not write unretrieved content as literature conclusions.

Manuscript Style

The manuscript is written as a standalone academic narrative, not as assignment instructions. When the starting object is a paper, use the title "Research Lineage". When the starting object is an author, use the title "Overview of Works". Section names use Chinese, not Intro or Background.

Fixed outline:

1. Introduction: which paper or author this article starts from, what this article answers, and how the following text is organized by theme.
2. Literature Sources: search route (Autoverse literature search), time limits, number of papers included, and abstracts serving as the basis for judgment. Do not write command names, and do not write the process of identifying papers or people.
3. Thematic Sections: when the starting object is a paper, organize by references, subsequent citations, or research questions. When the starting object is an author, organize by that author's research threads. Do not list paper by paper. Do not apply "points of agreement / disagreements and limitations" to every section.
4. Discussion: focus on agreement, disagreement, and insufficient evidence among studies.
5. Conclusion
6. References

The main text uses sequential numbering: mark after claims with [1], numbered consecutively by first appearance, with multiple papers written as [1, 4, 7]. The references at the end correspond one-to-one with the numbers. Row n of the evidence table is [n]. The order of the three places must be consistent. Do not write doi:… or arxiv:… strings in the main text.

For references, write as far as possible: author, title, conference or journal, year, DOI or URL. For preprints without formal publication, write the arXiv number. If the author is missing, start from the author field; leave that field blank.

The main text of the manuscript must not contain: delivery, agreement, user requirement, investigation, evidence table, identifier, not returned, checked items,resolve、batch、authors, Autoverse commands, split research questions. Write agreements and disagreements in the discussion, not as subsection headings.

Write English terms according to academic Chinese, do not translate them stiffly according to the original English words. For example: references, not "upstream"; subsequent citations, not "downstream"; preprint, not "preprint".

Use the relationship diagram only when it truly can explain the references or subsequent citations between two papers. When the content is dense, the Agent splits the relationship diagram into several.

Submission

The Agent explains the limits of the conclusions and the basis in the dialogue, and submits the following two items. The storage location is determined by the user. If the user does not specify, the Agent writes the files in the current working directory and informs the path in the dialogue. The Agent does not submit a PDF.

• Manuscript (.md): according to "Manuscript style".
• Evidence table (.csv): one row per paper, column order same as the reference numbering; list title, author, year, conference or journal, DOI or URL, document type, citation count, and points relevant to this question. Leave missing items blank.

The final draft writes only the research content and the limits of the basis.