Research To Paper Write

Write — draft and revise (self-contained, orchestrates audit + humanize)

This is the writing engine. It owns the parts unique to drafting — understanding the content, planning the rationale, and writing prose hedged to its evidence — and hands the two reusable passes to companion sub-skills: review to research-to-paper-audit, de-AI to research-to-paper-humanize. Splitting them out means each can also be used alone, but here they are run in sequence.

The discipline rests on one idea: good academic writing is a checklist applied in rounds, with the user in the loop. Never draft before understanding, never finish after one pass, never hand over prose that still reads as machine-generated. This is adversarial collaboration, not one-shot generation.

Two entry modes

The engine starts from one of two places — name which at the outset:

• Rewrite Existing — the user brings an existing paper/report/draft. Understand it, then strengthen it through

the audit (argument, hedging, structure, citations). Do NOT downgrade the task to surface polish — the audit explicitly rejects shallow/append-only edits.

• Build From Materials — the user brings a materials folder (notes, data/results, figures, PDFs, partial drafts,

experiment descriptions). Understand the materials, plan the rationale matrix, and draft from scratch.

Both modes then run the same Workflow below (understand → rationale → draft → audit → de-AI → build).

Workflow (run every time, in order)

1. Understand first — do not draft yet.

Read the user's materials: their data/results, any draft, the scope_brief.md (angle, target journal + limits, word count, themes) and the curated reference library if they exist. Note the entry mode (Rewrite Existing vs Build From Materials) and read the scene, research depth, and output language from the scope_brief.md. Restate, in your own words, the single core argument and the 3-5 supporting claims, and confirm them with the user. A beautiful paragraph built on the wrong claim is wasted; if you cannot restate the argument crisply, read more or ask before continuing.

1. Lock the controlling motivation, then plan the 论证脉络矩阵 (argument-thread matrix) — not a generic template.

First state the paper's one-sentence red thread and split it both ways — 要论证 (the prioritized claims) and 禁止论证 (what to NOT claim; the overclaim guardrails). Then split the work into real units (a claim, an evidence block, a synthesis move, a heading, a caption) and write one row each in writing_rationale_matrix.md, every row physically linked to the red thread through three non-empty anchors — 动机链接, 证据/引用锚, 文字动作 — plus its evidence-matched 措辞强度. Row F1 justifies the whole-work framing. Before drafting, run the offline gate python scripts/check_argument_matrix.py writing_rationale_matrix.md (pure stdlib, no network — works on a plain DeepSeek / no-VPN setup); it rejects empty anchors, generic-phrase reasons, a thin matrix, and a missing 禁止论证 side. Do not draft until it passes. See references/writing-craft.md §2 for the column spec and the bar.

1. Draft section by section, with evidence-matched hedging.

Write each unit to its matrix row. Match every verb to the strength of its evidence, and keep interpretation in Discussion, not Results — see references/writing-craft.md for the hedging ladder and the per-scene structures. The non-negotiable: a prediction stops at "predicted to / consistent with"; only a measured whole-system result may "confirm".

1. Review — hand to research-to-paper-audit.

Do not self-approve. Run the audit sub-skill: three independent reviewers (claim-support, logic, citation integrity) plus Editor Synthesis, re-run until a round is clean. Pass the scene (journal/conference/report/review/ competition, from the user or the scope_brief.md per "Target scenes") to the audit so reviewers apply scene-appropriate structural criteria. Apply its fixes before moving on.

1. De-AI — hand to research-to-paper-humanize.

Run the humanize sub-skill on the reviewed draft: the five-dimension pass (长短句 / paragraph variety / density / connectors / term variation), a recorded humanize_matrix, and the quantitative check. Iterate until it passes.

1. Discuss with the user at each round.

After understanding (step 1), after each audit round (step 4), and after de-AI (step 5), show what changed and what was flagged, and let the user steer. The user has the final say on every claim and edit — this is the "多轮对抗审查" the workflow promises.

1. Render to output formats. Once the user approves the final draft, hand it to research-to-paper-build to

produce LaTeX (.tex), Word (.docx), and optionally PDF, resolving citations from the library.bib.

Target scenes

The same engine produces several scenes; the structure differs (details in references/writing-craft.md). Pick the scene from the user or the scope_brief.md, and apply that venue's length and structure limits:

• journal / 论文 — IMRaD with strict Results-vs-Discussion separation, structured abstract, reproducible Methods.
• conference / 会议论文 — tighter and page/length-limited; the contribution stated up front, a condensed Methods.
• report / 报告 — problem → approach → findings → implications; plainer, audience-facing register.
• review / 综述 — argument-driven synthesis organized by theme, not by paper; each section advances a position.
• competition / 竞赛 — the novel contribution, results, and impact foregrounded against the competition's rubric.

Two more knobs come from the scope_brief.md:

• Research depth — flash (3 scene exemplars + 3 recent in-field papers + the venue's requirements) or pro

(6 + 6); deeper orientation reading and exemplar learning at pro.

• Output language — English or Chinese; draft in it, and run humanize_check.py --lang to match.

Files

• references/writing-craft.md — the hedging ladder, the per-scene section structures, and the 论证脉络矩阵 format (§2).
• scripts/check_argument_matrix.py — offline gate for the 论证脉络矩阵 (controlling-motivation doublet + per-row

anchors + no generic reasons + row-count floor); pure stdlib, runs with no network (DeepSeek / no-VPN friendly).

• Review and de-AI live in their own sub-skills: research-to-paper-audit and research-to-paper-humanize.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Jason-0409-G
• Source: Jason-0409-G/scriptorium
• License: MIT

Install and usage instructions live in the source repository linked above.