Think Before Write

A structured coarse-to-fine workflow for research paper development with AI assistance.

Five integrated Cursor skills that take a research paper from idea to submission-ready manuscript, with mandatory human checkpoints at every layer.

Skills

Skill	Purpose	Key Features
paper-storyline	Coarse-to-fine alignment (L0-L5)	6-layer hierarchy, human checkpoints, state tracking
academic-writing	Prose execution & QC	Track A/B (greenfield/brownfield), paragraph archetypes, Numbers Triangle, de-AI
paper-review	Simulated PC review	3 reviewer personas + PE + Meta-PC, Tier A/B rubric, Issue Cards, submission gate
author-supervisor	Strategic advice	5 independent professor lenses (Systems, Methodology, Writing, Strategy, Pragmatics)
paper-writing-router	Activation routing	Decision tree for which skill handles what

Plus:

• best-of-N-protocol: Reusable pattern for launching N parallel subagents
• Reference files: Writing patterns, sentence banks, anti-patterns, eval methodology, reviewer archetypes, rebuttal strategies, framing strategies (sourced from top-tier systems/compiler/HPC venues)
• Scripts: check-numbers.py, claim_audit.py, score_aggregate.py, review_preflight.sh

Architecture

paper-storyline (L0-L3)     Confirm contributions → sections → subsections → paragraphs
        │
        ▼
academic-writing (L4-L5)    Writing plan → prose execution → QC gates
        │
        ▼
paper-review                Simulated 3-reviewer PC → Issue Cards → submission gate

author-supervisor           Embedded advisory (any layer) — 5 professors in parallel
best-of-N-protocol          Reusable subagent ensemble pattern
paper-writing-router        "Which skill handles this?" decision tree

Installation

Copy the skills into your project's .cursor/skills/ directory:

# Clone
git clone https://github.com/SyntaxArchmage/think-before-write.git

# Copy all skills to your project
cp -r think-before-write/paper-storyline   your-project/.cursor/skills/
cp -r think-before-write/academic-writing  your-project/.cursor/skills/
cp -r think-before-write/paper-review      your-project/.cursor/skills/
cp -r think-before-write/author-supervisor your-project/.cursor/skills/
cp think-before-write/best-of-N-protocol.md your-project/.cursor/skills/
cp think-before-write/paper-writing-router.md your-project/.cursor/skills/

Usage

1. Start with storyline: Activate paper-storyline to align your paper's structure layer by layer
2. Write prose: Once L3 is confirmed, academic-writing handles Track A (new) or Track B (existing draft)
3. Review before submission: paper-review simulates a 3-reviewer PC with quantified scoring
4. Get strategic advice: Invoke author-supervisor at any point for 5 independent professor perspectives

The paper-writing-router resolves ambiguous requests (e.g., "review my intro" → academic-writing QC or paper-review simulation?).

Workflow

User: "Help me write this paper"
  → Router: structure question? → paper-storyline
  → Router: prose/QC? → academic-writing
  → Router: simulated review? → paper-review
  → Router: strategic advice? → author-supervisor (5 professors)

Best-of-N Subagent Modes

Several workflow steps support launching N parallel subagents for diverse output:

Mode	N	Skill
Abstract writing	5	academic-writing
L5 paragraph drafting	3	academic-writing
L3 storyline proposals	5	paper-storyline
De-AI polish	3	academic-writing
Review persona scoring	3×3	paper-review
Rebuttal drafting	5	paper-review
Author supervisor	5	author-supervisor

See best-of-N-protocol.md for the reusable pattern and session budget rules.

Scripts

Script	Purpose
check-numbers.py	Verify numeric claims against manifest.yaml
claim_audit.py	Flag overclaims, undefined terms, abstract/eval mismatches
score_aggregate.py	Aggregate reviewer scores with persona-specific offsets
review_preflight.sh	Pre-review gate (compile, pages, numbers, claims, anonymization)

Verified

This skill system was verified through 2 rounds of 5-agent parallel review:

• Round 1: 17 issues found and fixed
• Round 2: 10 issues found and fixed
• All fixes verified

License

MIT