科研 Skill

文献调研语料库维护

literature-survey

Finds, verifies, and maintains a literature corpus from arXiv, OpenAlex, Crossref, and related databases with BibTeX entries and per-paper notes. Use when sourcing references for a paper.

发布者 Codar安全审核通过v0.1.0更新于 2026-11-18

SKILL.md

Literature Survey

When to load this skill

Load this skill when the user is:

  • looking for related work, prior art, or papers on a topic (Step 0)
  • starting a new research paper that needs a literature review
  • adding new references to an existing paper
  • migrating an unverified BibTeX file to a verified one
  • doing a deep-read pass on a set of references and producing per-paper notes
  • needing a reproducible record of which references have been verified vs which are still placeholders

Overview

A research paper's bibliography drives credibility. Stale, mis-attributed, or unverified entries embarrass the author at submission time. This skill defines a workflow of five steps plus an optional Step 0 (finding candidates) for building a verified literature corpus and a deliverables structure that future-you (or a collaborator) can audit and extend.

The steps:

  1. (When candidates are not yet in hand) Find them: see "Step 0" below.
  2. Add a verified BibTeX entry to references/bibliography.bib.
  3. Drop the published-manuscript PDF into references/pdf/<citekey>.pdf.
  4. Generate a text extraction for AI consumption.
  5. Write a survey note notes/survey_<citekey>.md.
  6. Append a row to references/_collection_log.md.

Each step is described in detail below, plus the directory layout, the survey-note template, and the collection-log template.

Directory layout

<paper-repo>/
  references/
    bibliography.bib          -- BibTeX (tracked); every entry verified
    _collection_log.md        -- per-paper verification status + corrections + notes
    pdf/                      -- published-manuscript PDFs (gitignored)
      <citekey>.pdf
      <citekey>-supp.pdf      -- supplementary material where applicable
      .txt/                   -- pdftotext -layout extractions (gitignored)
        <citekey>.txt
  notes/
    survey_<citekey>.md       -- per-paper survey note (~30-50 lines)
    README.md                 -- index of survey notes (by section, by affinity)

The references/pdf/ directory and its .txt/ cache should be gitignored; the bibtex + collection log + survey notes are tracked.

Step 0: Finding candidate papers

Skip this step when the user hands you the papers. Otherwise, before Step 1, load references/literature-search.md from this skill. It covers:

  • choosing a strategy (snowball / systematic / exploratory), and when a "no prior work does X" claim needs the systematic one;
  • where to search for scientific computing (arXiv math.NA and friends, OpenAlex, Crossref, zbMATH, SIAM, OSTI, Semantic Scholar, Google Scholar);
  • programmatic lookups (arXiv, OpenAlex, Crossref, OSTI APIs) and DOI-to-BibTeX as the hand-off to Step 1;
  • triage (must read / should read / cite only / drop) and the search-log table that goes in references/_collection_log.md.

Each kept candidate enters Step 1 with a DOI or arXiv ID. A DOI-resolved BibTeX record is a starting point, not a verified entry.

Step 1: Verified BibTeX entry

Add the entry to references/bibliography.bib. Required fields: must include venue + page numbers + DOI or arXiv ID. Reject placeholder entries (e.g. "Author, A. (year). Title. Some Venue. [VERIFY]"); these go in a separate "pending" section of the bib file or are tagged with note = {pending verification}.

Verification criteria for an entry to be considered "verified":

  • author list matches the published manuscript (not arXiv preprint if the two differ)
  • year is the published year (not arXiv submission year), unless the entry is genuinely arXiv-only
  • venue + volume + issue + pages all match the published version
  • DOI or arXiv ID is correct (verify by webfetch of the publisher page or the arXiv abstract page)
  • citekey follows project convention (typically lastname + year + short-tag, e.g. foster2021dad)

When a verified entry corrects an earlier unverified entry, document the correction in the collection log (Step 5). Do NOT silently overwrite.

Step 2: PDF acquisition

Drop the published-manuscript PDF into references/pdf/<citekey>.pdf, where <citekey> matches the BibTeX citekey exactly. Supplementary material goes as <citekey>-supp.pdf.

The PDFs are gitignored (they're typically large; copyright varies; not appropriate for tracked content). The references/pdf/ directory itself should also be gitignored (or the directory tracked but the *.pdf glob ignored, depending on your project's convention -- see the .gitignore template in templates/paper-skeleton/).

If the published PDF is paywalled and only an arXiv preprint is available, use the preprint and explicitly document this in the collection log so a future reader knows to upgrade when access becomes available.

Step 3: Text extraction

Most AI agents cannot consume PDF binaries directly. Generate a layout-preserving text extraction once, then read it whenever a deep dive is needed:

mkdir -p references/pdf/.txt
pdftotext -layout references/pdf/<citekey>.pdf references/pdf/.txt/<citekey>.txt

pdftotext ships with poppler-utils (Linux: apt install poppler-utils; macOS: brew install poppler). The -layout flag preserves the two-column structure of typical academic papers, which makes the output much more readable than the default flow-mode extraction.

For supplementary material, repeat with <citekey>-supp.pdf -> .txt/<citekey>-supp.txt.

The .txt/ directory is gitignored; it is a regenerable cache.

Batch conversion of all PDFs in one go:

mkdir -p references/pdf/.txt
for pdf in references/pdf/*.pdf; do
  base=$(basename "$pdf" .pdf)
  pdftotext -layout "$pdf" "references/pdf/.txt/${base}.txt"
done

Step 4: Per-paper survey note

Read the .txt extraction, then write notes/survey_<citekey>.md to ~30-50 lines covering the structure in references/survey-note-template.md (loaded on demand from this skill's references/ subfolder when needed).

The survey note is the primary artefact of this workflow: it is what future-you reads when drafting the paper's related-work section, and it is what a collaborator reads to get up to speed without re-reading every PDF.

Briefly, a good survey note has these sections:

  1. Header: citekey, full citation, PDF path.
  2. Headline claim (one sentence).
  3. Method (full detail with key equations in MathJax).
  4. Test cases + parameters used in the paper.
  5. Headline numerical results.
  6. Relevance to our paper (why we cite this; in which sections).
  7. Critical observations (insights NOT in the abstract).
  8. Action items for our paper draft.

Load references/survey-note-template.md from this skill for the full template + a worked example.

Step 5: Collection log

Append a row to references/_collection_log.md documenting:

  • citekey
  • verification status (verified / arXiv-only / placeholder)
  • date verified
  • any corrections made vs an earlier entry
  • any quirks worth noting (e.g. "supplementary material is essential", "v3 of the arXiv preprint is significantly different from v1")

The collection log serves as the audit trail. When a reviewer asks "how did you decide on this citation?" or when you discover a mis-attribution months later, the log is where you check.

Load references/collection-log-template.md from this skill for the template.

Workflow rules

  1. Never cite a paper you have not verified. Placeholder entries are acceptable in references/bibliography.bib only if they are tagged explicitly (e.g. note = {pending verification}) and marked in the collection log as such.
  2. Survey-note quality matters more than count. A short paper with 5 well-written survey notes is more useful than a long one with 25 superficial ones. If a note would just restate the abstract, skip it and link the abstract instead.
  3. Re-read the source. When writing the paper's related-work section, re-open the survey note (and if needed the .txt extraction) rather than relying on memory or your initial summary.
  4. Track corrections. If you discover a mis-attribution (wrong author list, wrong year, wrong arXiv ID), log it explicitly in _collection_log.md. Do NOT silently fix it -- the audit trail is what makes the corpus trustworthy.
  5. MathJax for equations. Survey notes use MathJax ($...$ inline, $$...$$ display) for any equations transcribed from the source. ASCII-art math is forbidden in survey notes.
  6. Survey notes and the collection log are HUMAN-FACING documents. Both are read by the user when drafting the paper, by co-authors when re-orienting, and (sometimes) by reviewers. They follow the conventions in the human-facing-doc-authoring skill (audience split, narrative prose over telegraphic fragments, tables where they aid scanning, date-stamps, no personal-path leaks). When producing or substantially revising either artefact, also load the human-facing-doc-authoring skill (if installed) for the universal conventions and the per-doc-type self-review checklist.

Output contract

When the user invokes this skill, the agent should:

  1. Confirm or create the directory layout (Section "Directory layout") if it does not exist.
  2. Walk through the 5 steps for each paper the user provides (interactively or in batch).
  3. After processing, summarise in a single message:
    • Citekeys added.
    • Verification status of each.
    • Any corrections discovered vs prior entries.
    • PDFs missing (if any).
    • Survey notes written.
    • Suggested next step (typically: read the survey notes, then update the paper's plan-of-record / Section 1 reading list).

See also

  • references/literature-search.md -- Step 0: search strategies, sources, programmatic lookups, triage, and the search log. Adapted from genesis-skills literature-search (Apache-2.0).
  • references/survey-note-template.md -- the canonical survey-note template + worked example.
  • references/collection-log-template.md -- the canonical collection-log template + worked example.
  • templates/paper-skeleton/ (in the upstream scicomp-research-skills repository this skill was adapted from, sibling of skills/) -- a starter paper-repo skeleton that pre-creates the references/ and notes/ directory structure; not bundled with this skill, so create the directories manually if this template is unavailable.
  • The human-facing-doc-authoring skill (if installed) -- universal conventions for human-facing docs. The survey-note + collection-log templates above embody these conventions; load the human-facing-doc-authoring skill when authoring or revising either artefact and you want the cross-cutting checklist.
  • The agent-resource-discipline skill (if installed) -- a literature-survey pass typically processes many PDFs and iterates across sessions, so it is one of the heaviest token consumers in this ecosystem. Load this skill (and in particular its agent-resource-discipline/references/pdf-lifecycle.md and agent-resource-discipline/references/persistent-memory.md) at the start of any non-trivial literature pass: it codifies the one-shot pdftotext + survey-note-first lookup pattern, the context-window budget for handling many references, and the first-/ last-action index-update protocol that makes a multi-session pass cheap.

Created 2026-05-13 by A. Attia. Distilled from the literature-survey workflow developed for a research paper (14 references verified across 3 sections of the paper's plan-of-record). Revised 2026-05-13 (added Workflow rule #6 + See-also pointer to human-facing-doc-authoring skill, since survey notes and the collection log are human-facing artefacts). Revised 2026-05-13 (added See-also pointer to agent-resource-discipline skill, since a literature-survey pass is PDF-heavy + multi-session and therefore benefits explicitly from the PDF-lifecycle + persistent-memory + context-window-budget protocols that skill codifies). Revised 2026-10-14 (added Step 0 "Finding candidate papers" + references/literature-search.md, adapted from the Apache-2.0 genesis-skills literature-search skill by I. Foster; description extended to cover finding papers).

示例提示词

帮我从多个学术数据库搜集候选文献,整理成带笔记的调研语料库。

所属插件

跨学科科研方法跨学科通用科研方法:系统综述与元分析、预注册、实验与问卷设计、文献矛盾与研究空白、科学计量与专利分析、可复现研究、答辩与审稿回复。

安装该插件会一并获得同领域的 54 个 Skill / MCP。

同插件的其他科研 Skill

常见问题

文献调研语料库维护能做什么?

Finds, verifies, and maintains a literature corpus from arXiv, OpenAlex, Crossref, and related databases with BibTeX entries and per-paper notes. Use when sourcing references for a paper。

如何安装文献调研语料库维护?

在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「文献调研语料库维护」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。

文献调研语料库维护需要哪些权限?

该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。

文献调研语料库维护经过官方审核吗?

是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。