researchwrite

Original：🇨🇳 Chinese

Translated

1 scriptsChecked / no sensitive code detected

Proposal-first scientific writing pipeline. Three modes (compose/revise/hybrid) with a four-layer QA pipeline. It enforces the principles of evidence-before-prose, argument-before-sections, and contracts-before-paragraphs.

11installs

Sourceyuan1z0825/nature-skills

Added on2026-07-16

NPX Install

npx skill4agent add yuan1z0825/nature-skills researchwrite

SKILL.md Content (Chinese)

View Translation Comparison →

researchwrite — Proposal-first Scientific Writing Pipeline

A scientific writing state machine inspired by autonovel (state machine + scoring), professor (dynamic experts), brainstorming (entry-level questioning), and anti-AI-writing (language cleaning). Not a general "help me write a paper" prompt.

Core Principles

Evidence Before Prose — Must establish or read
```
research_canon
```
and
```
evidence_table
```
before drafting
Argument Before Sections — Must complete
```
argument_map
```
before writing the main text
Contracts Before Paragraphs — Each section requires purpose / allowed claims / forbidden claims / inputs / validation
Scope Before Completeness — If writing in phases, first lock the phase boundaries
Dynamic Experts, No Fixed Pool — Use
```
professor
```
to summon corresponding experts based on failure modes
Content Before Language — Diagnose scientific logic before performing anti-slop / language polishing
No Automatic Fact Upgrading — Never change "may indicate" to "proves" unless supported by evidence
Deletion Over Explanation — When a claim is not feasible, delete it directly. Keep the main text clean; leave explanations for defense
Know When to Stop — Plateau periods, expert conflicts, and missing evidence are reasons to stop, not to polish

Mode Assignment

Input	Mode
Topic, direction, vague ideas	`compose` — Load `references/compose-mode.md`
Existing paragraphs/sections/complete proposal	`revise` — Load `references/revise-mode.md`
Existing draft + expansion/supplementation/restructuring	`hybrid` — Load `references/hybrid-mode.md`

Infer the default mode when ambiguous. Only ask the user if the choice will change the workflow.

Project Structure

Working Directory:

<outputs>/researchwrite/<project-slug>/

Standard Files:

00_scope.md              Writing task boundaries
01_research_canon.md     Hard facts and constraints
02_evidence_table.md      claim → evidence mapping table
03_argument_map.md        Argument framework
04_section_contracts.md   Purpose / inputs / allowed & forbidden claims for each section
05_style_guide.md         Style, terminology, prohibited expressions
state.json                Project status (mode, round, scores, technical_debts)
sources/                  User materials, literature, data
drafts/                   Drafts and section files
revision_briefs/          Revision briefs
qa_logs/                  Diagnosis, expert reviews, anti-slop, scoring records
exports/                  Final outputs (.md + .docx)

When creating a new project, take empty templates from

templates/

.

references/worked-example-quaternary-proposal.md

provides a fully filled example based on the general materials science field.

Reference File Index

The public version includes four key references:

Reference	Purpose
`references/evaluation-rubric.md`	8-dimension × 4-anchor scoring system + scoring process
`references/stopping-rules.md`	Iteration loop termination conditions
`references/foundation-files.md`	Establish the five foundation files
`references/validation-checklist.md`	Automatic validation checklist

The complete reference files (12 documents including detailed explanations of compose/revise/hybrid modes, expert assignment, export archiving, review framework, etc.) are not included in the public version. To obtain them, please contact the author via GitHub issue or email.

Running Delivery

At the end of each run, output:

Current file path or revised text
Current score/status
Remaining risks
A suggested next step

QA Mode — Four-layer Quality Assurance Pipeline

Enter QA mode when the user says "review this section/run QA/check the plan/go through the pipeline".

Scenario Gears

Gear	Applicable Scenario	Threshold
`paper`	Submitted paper	7.0
`proposal`	Research proposal/thesis proposal	7.0
`internal`	Internal report/weekly report	5.0
`quick`	Quick scan	None

Pipeline Order (Content Before Language)

Gate 2: Professor Convener (Content Layer)
  ├── Paper → Methodology expert + Domain expert
  ├── Proposal → Feasibility expert + Innovation expert
  └── Literature reviews → Coverage expert + Critical depth expert
  │
  ▼
Gate 1: avoid-ai-writing mode (detect-only)
  └── Only valid for English; skip or downgrade to manual scanning for Chinese
  │
  ▼
Gate 3: Auto-validation (Format/Completeness Layer)
  └── Does each claim have a citation? Is the method reproducible? Are numbers consecutive?
  │
  ▼
Gate 4: Scoring Threshold (Dimension-based Scoring)
  ├── Total score ≥ threshold → Pass ✅
  └── Total score < threshold → Directed rollback based on low-score dimensions (no more than 3 rounds)

Gate 2 comes before Gate 1 — Avoid rewriting sentences only to be sent back by experts.

Usage Examples

User: "Review this discussion section with researchwrite, paper gear"
  → Automatically go through Gate 2 → Gate 1 → Gate 3 → Gate 4
  → Return review report + targeted modification suggestions

User: "Quickly scan this email"
  → Use quick gear, only mark P0 issues

Configure Your Research Domain

Tell the agent your research background on first use, and the agent will call

professor

to establish domain expert knowledge. Subsequent expert reviews during writing will be based on your field.