Research & Analysis Workflows

Umbrella, fully self-contained workflow that takes a research idea all the way to a submission-ready manuscript, depending on NO other skill. It chains stages you can also run on their own: SCOPE (research the field, discuss directions one question at a time, lock angle/target journal/word count), CURATE (build a DOI-verified library — RIS for Zotero/EndNote + by-category Excel — behind a multi-agent review), WRITE (draft, then review and de-AI), and BUILD (render to LaTeX/Word/PDF). WRITE is backed by an AUDIT step (independent multi-reviewer rounds) and a HUMANIZE step (five-dimension de-AI incl. 长短句). Use whenever a request spans more than one stage, for example 帮我从定方向到写出来, 做文献综述, 整理文献库并核对DOI再写论文, 从一个想法做到初稿, 我想投某期刊先定方向再建库再写, "take this idea to a written, formatted paper end to end". Drafting one piece alone (帮我写论文/综述/报告) goes to research-to-paper-write, not here. For a single piece, use the sub-skill directly — angle/journal → research-to-paper-scope; verify references / export a library → research-to-paper-curate; draft or rewrite → research-to-paper-write; review/audit a draft → research-to-paper-audit; de-AI / 降AI率 → research-to-paper-humanize; export to LaTeX/Word/PDF → research-to-paper-build. Checking whether a run's artifacts are complete, or whether a draft's [@key] / \cite{} citations resolve against library.bib (just checking that they resolve, NOT rendering them) → research-to-paper's own artifactcheck / citecheck scripts.

Adversarially review an academic draft with independent multi-reviewer rounds, depending on no other skill. It spawns three independent reviewer agents — claim-support, logic/structure, and citation/evidence integrity — each reading only its own prompt and the draft, then writes an Editor Synthesis and re-runs until a round comes back clean. It catches unsupported claims, claims drifted from their evidence tier, interpretation leaking into Results, shallow/append-only revisions, and citations that do not actually support the sentence. Use whenever the user has an EXISTING draft to peer-review or stress-test — this skill reviews a draft, it does NOT write a 综述/review article; if the user has no draft yet and wants a review article written, use research-to-paper-write. Triggers: the user says 审一下这篇草稿, 多轮审, 帮我挑毛病, 查无据claim, 逻辑审一遍, 这篇有没有过度声称, "review my draft", "audit this manuscript", "check for unsupported claims", "is anything overclaimed", "stress-test the argument". Works on a whole paper or a section; called by research-to-paper-write after drafting, or invoked directly on any draft.

Render a finished manuscript into submission formats — LaTeX (.tex), Word (.docx), and optionally PDF — depending on no other skill. It runs pandoc on the manuscript markdown, resolving [@citekey] citations from a library.bib into a formatted reference list, and can inherit a Word template's styles or compile a PDF when a TeX engine is present. Use whenever the user wants the writeup in a specific output format, says 导出成LaTeX, 生成Word, 转成docx, 排版成论文/PDF, 输出tex, 这个稿子转成latex和word, "export to LaTeX", "give me a Word doc", "render to PDF", "build the .tex". Runs the bundled scripts/buildformats.py; takes the manuscript from research-to-paper-write and the library.bib from research-to-paper-curate.

Build a trustworthy literature library with NO dependency on any other skill: search five literature databases (OpenAlex, Europe PMC, PubMed, Semantic Scholar, CrossRef), verify every DOI against CrossRef (catching dead DOIs and the worse case of a DOI that resolves to the wrong paper), run a multi-agent adversarial review to reject fabricated or mis-attributed references, then either import straight into Zotero (via its Web API, one collection per category, when ZOTEROAPIKEY is set) or export a single RIS that imports into both Zotero and EndNote, plus BibTeX and a rich color-coded Excel that gives every paper a one-line 中文重点 (Chinese key point) alongside category, DOI-verification status, and optional innovation score. It also searches biology RESOURCE databases — any NCBI Entrez db (protein, nucleotide, gene, taxonomy, assembly, structure, SRA, BioProject, ...) plus UniProt, RCSB PDB, AlphaFold and Europe PMC. Use whenever the user wants to collect, check, search, or export references or bio records, e.g. 整理文献库, 把这批参考文献核对一下DOI, 核对引用是否真实, 导成能进EndNote或Zotero的库, 直接导入Zotero, 给文献配个分类Excel, 把每篇文章的重点用中文说明, 整理成带中文重点的Excel, 搜NCBI/蛋白/序列/结构数据库, "verify these DOIs", "check these citations are real", "export these papers to EndNote", "search UniProt/PDB/NCBI", "build me a reference library". Runs the bundled scripts searchpapers.py, verifydoi.py, exportrefs.py, biosearch.py and pushzotero.py; takes optional category themes from a scopebrief.md.

Reduce the AI-detection signature of academic prose and make it read like careful human writing, with no dependency on any other skill. It works across five dimensions — D1 sentence-length variation (长短句结合), D2 paragraph-structure variety, D3 information-density alternation, D4 connector control (removing 首先/其次/ 此外/值得注意的是 and other AI-frequent connectors), D5 term-context variation — at a light/medium/heavy tier, records every change in a humanizematrix.md, and verifies the result with a quantitative checker. Use whenever the user wants to de-AI text or lower an AI-detection rate, says 降AI率, 去AI腔, 去除AI痕迹, 这段太像AI改一下, 长短句结合, 学术润色降AI, "reduce AI detection", "humanize this draft", "make this sound less like ChatGPT", "vary the sentence length". Works on a whole paper or a single paragraph; called by research-to-paper-write as its final pass, or invoked directly on any text.

Decide the direction of a paper BEFORE searching or writing, by researching the field first and then discussing with the user one question at a time. Given a broad topic, it reads a few authoritative sources (recent high-impact reviews + seminal works), then proposes 2-3 concrete angles with trade-offs, and locks the entry mode (build-from-materials vs rewrite), angle, scope, target scene (journal/conference/report/review/ competition), target journal, research depth (flash/pro), word count, output language (中文/English), and core themes through step-by-step confirmation. Once the target journal is set, it looks up that journal's author requirements online. Use whenever the user says 帮我定研究方向, 我有个大方向想缩成题目, 这个综述/论文该从哪个角度写, 帮我选切入点, 我想投X期刊先确定方向和要求, "help me scope this", "narrow my topic", "what angle should I take", "figure out the direction before I start". Produces a scopebrief.md that feeds research-to-paper-curate (categories) and research-to-paper-write (motivation + journal limits).

Draft and revise an academic manuscript end to end in either of two modes — Build From Materials or Rewrite Existing — across scenes journal/论文, conference/会议论文, report/报告, review/综述, competition/竞赛, at a flash or pro research depth. It understands the content first, plans a per-unit writing rationale (not a generic IMRaD template), and drafts with evidence-matched hedging, then runs the review and de-AI passes through its companion sub-skills research-to-paper-audit (independent multi-reviewer rounds) and research-to-paper-humanize (five-dimension de-AI incl. 长短句), discussing with the user at each round. Depends on no skill outside this plugin. Use whenever the user wants to write or rewrite a paper, review, or report, says 帮我写论文/综述/报告, 把这些内容写成稿子, 改写这篇草稿, 按我的写作纪律来, 写会议论文/竞赛论文, 从素材构建, "draft this into a paper", "build a paper from these materials", "rewrite this manuscript", "write up these results". Reads an optional scopebrief.md (angle/journal/word-count) and a curated reference library if present; works standalone otherwise. For ONLY a review pass use research-to-paper-audit; for ONLY de-AI use research-to-paper-humanize.