universe-research

Universe Research

Help students complete literature retrieval and write a review on a computer science, medical, or engineering problem. The Agent is responsible for determining the discipline, formulating search queries, comparing literature, and writing the main text. Autoverse is responsible for retrieval and verifying bibliographic entries. The Agent does not write separate retrieval programs.

The execution surface for literature retrieval and verification is autoverse. Commands, parameters, and error codes are governed by ../autoverse-cli/references/commands.md . The Agent does not invent option names, must not directly request HTTP, and must not browse the OpenAPI on its own to choose paths.

For discovery and identifying papers, preferentially use the curated verbs:

autoverse --json --quiet search "<research question>" --domain <见对应 reference>
autoverse --json --quiet search "<query>" --ccf A --conference --year-from <year>
autoverse --json --quiet resolve <prefixed-id>
autoverse --json --quiet batch -f <本次任务标识文件>

When the curated verbs lack required fields, only the following shared paths may be used; for medicine, the two supplementary paths explicitly listed in the medical reference may also be used. Unlisted /v1 paths must not be called.

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

Discipline-specific search angles, aliases,search --domain, topic organization, and disciplinary terminology are written in references/computer-science.md、references/medicine.md and references/engineering.mdrespectively. The discipline reference only supplements the above content and does not rewrite the rules in this document regarding the recent five years, foundational works, title search, and verification. The Agent determines the discipline before starting retrieval and first reads the corresponding reference. The Agent must not apply the search angles of one discipline to another discipline. Medical supplementary paths are written only in the medical reference.

Writing and terminology: before writing the main text, read references/chinese-academic-prose.md and references/chinese-academic-terms.md. For citation keys and compilation, see references/citations.md。

Practices

• The Agent makes the first necessary call. When INVALID_SESSION is returned, the Agent tells the user to log in, and after logging in continues the original task. The Agent should not open the login flow on the user's behalf.
• The Agent first determines whether the problem belongs to computer science, medicine, or engineering. If the user has already stated the discipline, the Agent does not ask again. If the discipline cannot be determined from the problem, the Agent asks only one brief question. The Agent retrieves only after reading the corresponding reference.
• Topic retrieval focuses mainly on the recent five years, while foundational research on the topic is still included. Entries with empty abstracts must not be used as a basis for drawing conclusions. The Agent's deliverables are: a review (.md) and an evidence table (.csv), and no PDF is submitted.
• The Agent breaks the overall question apart according to the search angles in the discipline reference. Each angle is searched separately. When a broad query is too noisy, or when a known work needs to be found again, switch to a full title search without a year window. The Agent first deduplicates by DOI; when there is no DOI, works are considered the same only if the title, main author, and year all match. All aliases listed in the discipline reference must be included in the search queries. When a preprint and the formally published text of the same work appear side by side, the Agent keeps only the formally published version.
• During retrieval, combine discipline-authoritative catalog labels (such as CCF ranking, CAST catalog, Peking University Core) with literature type (review/journal/conference) to carefully screen candidate entries.
• The Agent uses Autoverse to retrieve multiple times. The Agent revises the search queries based on the results. The Agent must not substitute the number of calls or retrieval ranking for judgment about the literature.
• When searching for literature, the Agent first uses Autoverse retrieval. Web search and official pages such as those of publishers, conferences, PubMed, or arXiv may be used to supplement retrieval and to verify the bibliographic records already written on that page.
• When verifying bibliographic entries, the Agent writes the identifiers to be included in the review or evidence table into a task file for this task in the system temporary directory (the file name contains the task short name or date, not the fixed name ids.txt), using batch One verification. The Agent must not verify each article individually. A webpage cannot prove the reference relationship between two papers or subsequent citation relationships. DOIs, PMIDs, PMCIDs, or arXiv numbers written into the review or evidence table must be verified through batch verification. Articles that cannot be verified must not be treated by the Agent as cited references.batch Entries that succeed but whose title, authors, or abstract do not match the work to be cited shall be removed by the Agent.batch Successful packages must be saved as JSON for compiling citations.
• The Agent includes literature until the returned literature is sufficient to compare the conclusions of the studies and to state agreement, disagreement, or insufficient evidence. The Agent must not include literature unrelated to the topic merely to pad the number of articles. When retrieval or scoring is insufficient, the Agent shall state the actual coverage in the literature sources and must not write unretrieved content as literature conclusions.

Citations and Basis

Claims and literature points in the review must be traceable to verified articles with non-empty abstracts. An abstract supports only the judgments that the abstract can support. The Agent must not fabricate DOIs, PMIDs, arXiv numbers, titles, or research conclusions.

Short abstracts in bibliographic records are for screening only. Before writing, the Agent must read the full abstract and the returned fields. Authors, year, conference or journal are taken from the bibliographic records returned by Autoverse. If the search results give only the first author, the Agent uses batch to retrieve all authors. Citation counts and literature type: use them if returned. Leave still-missing fields blank. The Agent does not write "not returned." The Agent does not sort by citation count.

Articles with empty abstracts may be written into this task's identifier file and included in one batch。batch success allows entry into the evidence table and references; leave the key points column blank. The topic sections and discussion must not, based on that article, state what the work did, what it proved, or with whom it agrees or disagrees. The Agent must not call resolve or batch。

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

When the basis is insufficient or literature conclusions are inconsistent, the Agent shall state this in the discussion and must not smooth it over.

A medical review does not constitute individual diagnostic or treatment advice. The Agent states this only in the conversation. The Agent does not write this statement into the review title or opening.

Review Format

The review is written as a standalone academic review of that discipline, not as an assignment description. Use "Review" in the title, not "Narrative Review." When contrast with systematic reviews is needed, the Agent writes "narrative review" in the literature sources. For the writing style and the responsibilities of each section, see references/chinese-academic-prose.md. For the first appearance of proper nouns, see references/chinese-academic-terms.md。

Fixed outline, section names in Chinese, do not use Intro, Background:

1. Introduction: the object in the user's topic, and the scope compared in this article. Do not first define the entire discipline.
2. Literature Sources: search channels (Autoverse literature search; if official pages were used as supplements, state the sources as well), time limits, how aliases were included, number of articles included, and abstracts used as the basis for judgment. Do not write command names, and do not write the bibliographic record verification process. The disciplinary statements in the literature sources shall be based on the corresponding reference.
3. Topic Sections: organized by the topics in the corresponding reference. Do not list article by article. 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(written by the script; the Agent does not write this section)

Use standard citations in the draft body: after a claim, write [@DOI], multiple writing pieces [@DOI1; @DOI2], format see references/citations.md. The Agent does not hand-write [1], and does not hand-write the references at the end. After finishing, run:

python "<本技能包根目录>/scripts/render_cites.py" --json <batch成功包> --draft <综述草稿.md> --out <综述.md> --csv <证据表.csv>

The root directory of this skill package is the SKILL.md directory containing universe-research . When the Agent runs in the user's research directory, use that directory's absolute path, do not use the one under the current working directory scripts/。

The script numbers consecutively by first appearance. The submission review, the references at the end, and line n of the evidence table must be the output of that script, and the order must be consistent across all three places. The submission body text does not write doi:… or [@…]. The Agent only fills in the "key points" on the evidence table written by the script.

The reference citations are populated by the script from JSON: authors, title, conference or journal, year, DOI or URL. For preprints without formal publication and without a DOI, write the arXiv number. If authors are missing, start from the author field and leave that field blank.

The review body text must not contain: delivery, agreement, user request, survey, evidence table, identifier, not returned, checked items,resolve、batch, Autoverse command, split research questions. Agreements and disagreements are written into the discussion, not as subheadings of each section.

Discipline-specific forbidden literal translations are governed by the corresponding reference.

Submission

The Agent submits the following two items. The save location is subject to user specification. If the user does not specify, the Agent writes the files in the current working directory and tells the path in the conversation. The Agent does not submit a PDF.

• Review (.md）：render_cites.py the compiled manuscript, according to the "Review Style Guide".
• Evidence table (.csv): the same script writes it out with the same numbering; list title, authors, year, conference or journal, DOI or URL, document type, citation count, and key points relevant to this question. The first seven columns must not be manually modified. Leave missing items blank. For items with an empty abstract, leave the key points column blank.