---
name: research-to-paper
description: >-
  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 artifact_check / cite_check scripts.
---

# Research-to-Paper (orchestrator)

Entry point for a self-contained workflow. Nothing here calls another plugin — search, verification, review,
writing, de-AI, and format rendering all live in this skill's six sub-skills, so it works standalone (a
collaborator installs this one plugin and has everything). Find which stage the user is at, run the stages they
need in order, and pass each stage's output into the next.

## The stages

1. **SCOPE — decide the direction** → **`research-to-paper-scope`**.
   Reads authoritative literature, then confirms an angle, scope, **target journal**, and **word count** one step at
   a time, and looks up the journal's author requirements once chosen. Output: `scope_brief.md`. Skip only if the
   user already has a confirmed direction.

2. **CURATE — build a verified library** → **`research-to-paper-curate`**.
   Built-in five-source search, per-paper CrossRef DOI verification, a multi-agent adversarial review gate, then
   export to RIS (Zotero/EndNote) + BibTeX + a by-category Excel. The SCOPE themes become the categories.

3. **WRITE — draft, review, de-AI** → **`research-to-paper-write`**, which orchestrates two companion sub-skills:
   **`research-to-paper-audit`** (independent multi-reviewer rounds) and **`research-to-paper-humanize`**
   (five-dimension de-AI incl. 长短句). Understands the content first, plans a per-unit rationale, drafts with
   evidence-matched hedging, then audits until clean and de-AIs — discussing with the user each round. Scenes:
   `journal` (论文) / `conference` (会议论文) / `report` (报告) / `review` (综述) / `competition` (竞赛).

4. **BUILD — render to formats** → **`research-to-paper-build`**.
   Turns the approved manuscript into LaTeX (`.tex`), Word (`.docx`), and optionally PDF via pandoc, resolving
   citations from the `library.bib`.

## How to route

- Vague idea → SCOPE → CURATE → WRITE → BUILD.
- Has a direction, needs references → CURATE → WRITE → BUILD.
- Has direction + library, wants to write → WRITE → BUILD.
- "Verify these DOIs" / "export to EndNote" → CURATE.
- "Help me figure out the angle / which journal" → SCOPE.
- "Draft / rewrite this" → WRITE.
- "Review / audit this draft" / "查无据claim" → research-to-paper-audit.
- "降AI率 / de-AI this" / "长短句" → research-to-paper-humanize.
- "导出 LaTeX/Word/PDF" / "转成 docx" → research-to-paper-build.
- "产物齐了吗 / 检查产物 / verify the artifacts" → `scripts/artifact_check.py`.
- "引用对得上吗 / 引用能解析吗 / check the citations resolve" → `scripts/cite_check.py`.

Each stage is independently useful, so don't force the full pipeline when the user asked for one piece. Hand off
cleanly and tell the user what each produced before moving on.

## Workspace & artifacts

Run a multi-stage job inside one working directory so the *reasoning* trail is auditable, not just the final file.
Scaffold it once at the start:

```
python scripts/new_workspace.py manuscript_workspace
```

Each stage writes its artifact into this workspace — `scope_brief.md`, `library/`, `writing_rationale_matrix.md`,
`draft.md`, `audit_report.md`, `humanize_matrix.md`, `final/`. The full list (what each records) and the
verification commands are in `references/artifacts.md`. Before calling a run done, check completeness and that
every citation resolves:

```
python scripts/artifact_check.py manuscript_workspace --write   # refreshes artifact_manifest.md
python scripts/cite_check.py manuscript_workspace/draft.md manuscript_workspace/library/library.bib
```

A polished draft with an empty rationale matrix or no audit report is an incomplete run, not a finished one.

## Files

- `scripts/new_workspace.py` — scaffold the `manuscript_workspace/` and seed `scope_brief.md`.
- `scripts/artifact_check.py` — verify the artifact trail; write `artifact_manifest.md`.
- `scripts/cite_check.py` — check draft `[@key]` / `\cite{}` citations resolve against `library.bib`.
- `references/artifacts.md` — the workspace layout, the full artifact set, and every check command.