research-paper

name: research-paperdescription: >  Use when the user asks to write or draft an ORIGINAL RESEARCH ARTICLE —  IMRaD paper, conference paper, short/workshop paper, 研究论文/期刊论文/会议论文 —  reporting their own completed experiments, in English or Chinese. Triggers:  "write the paper", "draft our manuscript", "把我们这篇工作写成论文", "论文写作".  Results-first: the user's actual results (tables, metrics, figures) are the  raw material; this skill never invents an experiment or a number. NOT for  literature reviews/surveys (use medical-imaging-review), NOT for proposals  (use research-proposal), NOT for grant applications.

Research Paper Writer

Draft an original research article from the user's actual completed work: their experiments, data, figures, and references. Two disciplines govern everything:

1. Results-first. Methods and Results are written from materials the user supplied (result tables, metrics dumps, figures, code notebooks). Anything not supplied is flagged [PENDING: …] — never filled with a plausible number. A fabricated result is worse than a fabricated citation: it is data fraud, not a style slip.
2. Citation integrity (this repository's shared five-rule core, instantiated for author-year/numeric journal styles — see references/CITATION_INTEGRITY.md): every reference verified to exist before it enters the bibliography; [UNVERIFIED] flags for the rest; no invented DOIs, authors, or numbers.

Scenario routing

Read references/STRUCTURE_GUIDE.md for section-by-section construction and references/REPORTING_CHECKLISTS.md for what the binding checklist actually requires.

Phase 1: Intake (record, don't improvise)

Use AskUserQuestion to collect:

1. Study type + target venue/journal — routes structure and checklist (above).
2. Materials in hand — result tables/CSVs, metric numbers, figures, prior drafts/notes, code notebook. If the user has no results yet, stop: this skill writes papers about completed work, not imagined work (that is research-proposal territory).
3. References pool — user's Zotero/bib, or run the lit-search skill for a defensible related-work corpus (its references.bib drops straight in).
4. Language and length — en/zh; word limit from the venue's guide.

Record the intake as a short paper_plan.md (study type, venue, checklist, materials list, citation style). Headless runs: record inline and continue.

Phase 2: Materials audit (before any prose)

Inventory every number that will appear in Results:

• Each metric → its source artifact (file, table, cell). Untraceable numbers are [PENDING], and the user is told before drafting starts.
• Each figure → real file or a [PENDING: figure] placeholder with a content description. Redrawing/schematics follow the same honesty rules as scholar-slides (label what is not data).
• This inventory becomes results_ledger.csv (see references/QUALITY_CHECKLIST.md) — the paper-side sibling of the review skill's claim ledger: one row per load-bearing number, claim_id → section → claim_text → value（列名以 references/QUALITY_CHECKLIST.md 的 schema 表头为准）.

Phase 3: Outline + approval (red line)

Section skeleton per references/STRUCTURE_GUIDE.md, with per-section bullet content mapped to Phase 2 materials. Present the outline and wait for explicit approval (same rules as research-proposal: "你看着办" → state assumptions, re-present, proceed).

Phase 4: Write-with-verify (per section)

Draft section by section into a single working file paper_draft.md (the Phase 5 gates grep exactly that filename — a renamed draft makes them scan an empty file and report a false zero-hit); after each section, before starting the next:

1. Every citation in the section is from the verified pool (rules: references/CITATION_INTEGRITY.md), written as a stable pandoc citekey [@smith2024model] that has an entry in references.bib — never a hand-numbered [17] in the working draft. The venue's numeric/author-year style is rendered only at the end (pandoc paper_draft.md --citeproc --bibliography references.bib --csl <venue>.csl).
2. Every number in the section reconciles against results_ledger.csv.
3. Self-check against the AI-tell list and style rules (references/WRITING_STYLE.md) — the Phase 5 grep gates make these checkable, not aspirational.
4. Tense/register per section (Methods: past, what was done; Results: past, what was found — no interpretation; Discussion: present, what it means).

Phase 5: Gates + reporting checklist + submission pack

Mechanical gates (paste command output as evidence). Run the executable gate first — it covers all four mechanical checks in one pass and names every offending line:

python3 <skill_dir>/scripts/check_paper.py paper_draft.md \
    --ledger results_ledger.csv --bib references.bib
# exit 0 = no mechanical findings (not proof the paper is right); 1 = findings; 2 = bad input
# --allow-disclosed-flags : [PENDING…]/[UNVERIFIED…] already disclosed → warnings, not failures
# --allow-uncited KEY     : bib entry listed on purpose without in-text citation (venue norm)
# --skip-checklist        : drafting-stage run before checklist_<name>.md exists (warned)

It checks: (1) every metric-like number in the prose (decimals, percentages, n=) is covered by a results_ledger.csv value/dispersion; (2) [@citekey] ↔ references.bib both directions; (3) the placeholder and AI-tell patterns below (same regex, locked by tests); (4) every checklist_*.md row carries a disposition. Without Python, the two grep gates are the minimum:

# 1) Placeholder gate — zero hits (or disclosed [UNVERIFIED]/[PENDING])
grep -nE "xxx|XXXX|\[TBD\]|\[PENDING|\[UNVERIFIED[^]]*\]|10\.xxxx|doi:10\.[a-z]+/x" paper_draft.md

# 2) AI-tell gate — zero hits
grep -nEi "over the past ((few|several|two|three) )?(decades?|years)|in recent years,|it is worth noting|plays a (crucial|pivotal|vital) role|sheds light on|paves the way|ever-evolving|has garnered significant attention|delves into|a testament to|众所周知|近年来，?随着|日益受到(广泛)?关注" paper_draft.md

Then: work through the binding reporting checklist item by item (references/REPORTING_CHECKLISTS.md); assemble the submission pack — title page, abstract in the venue's format, keywords, cover letter, author contributions (CRediT) / conflicts (references/SUBMISSION_PACK.md).

Reference Files

Read only the file for the phase you are in.

learnings：复盘先于重写

修改 references 或 scaffold 之前先扫 learnings/；返工或发现 自相矛盾后追加一篇，日期命名（格式见 lit-search/learnings/）。

Error handling

• No results materials — stop and say so; suggest research-proposal if the work is still planned.
• A claimed number can't be traced — [PENDING] + explicit disclosure, never a placeholder that looks real.
• Reference can't be verified — [UNVERIFIED] + disclosure (never dropped silently, never fabricated).
• Venue requires a checklist this skill doesn't carry — say so, point to the official source, and record it as a manual gate in paper_plan.md.