Research writing

Research Writing

Treat research writing as claim-evidence-audience engineering. The goal is not to make a draft sound academic; it is to make the paper's claims, evidence, structure, venue expectations, and uncertainty boundaries legible to the intended research community.

Load As Needed

• For mode execution details, read references/workflows.md.
• For writing and audit rules, read references/rules.md.
• For output formats, read references/output-templates.md.
• For interpreting quick_scan.py, read references/quick-scan-coverage.md.
• For preset venue guidance, read one file from references/venue-profiles/.
• For mechanical English draft scanning, run scripts/quick_scan.py.
• For Semantic Scholar literature search support, run scripts/semantic_scholar.py.
• For stable project state, use scripts/workspace_state.py.
• For manuscript assembly and fallback PDF production, use scripts/assemble_manuscript.py, scripts/markdown_to_latex.py, and scripts/compile_latex.py.

Entry Principle

Diagnose first, then route. Keywords are shortcuts only when the user's intent and materials are already clear.

Check five dimensions:

Dimension	Determine
Paper stage	idea, active drafting, full draft, post-review revision, submission preparation
Current material	none, paragraph, section, full manuscript, reviewer comments
Main pain	gap, structure, argument, style, venue fit, reviewer response, format
Target venue	specified, unspecified, deadline-sensitive
Workspace	output directory and paper name

If context is ambiguous, ask 1-3 necessary questions. Do not send a long questionnaire.

Workspace Rule

Before executing any file-writing mode, determine a workspace. Conversation text is only a view; durable work is written to files.

• If the user provides a path, use it.
• If no path is provided, propose <current-working-directory>/paper-review/<paper-name>/ and wait for confirmation.
• If .workspace.yml exists, load it and avoid re-asking known state.
• Use scripts/workspace_state.py init, update-mode, and validate when possible.

Recommended workspace shape:

<workspace>/<paper-name>/
|-- .workspace.yml
|-- 01-research-positioning.md
|-- 02-writing-audit.md
|-- 02-writing-audit-v2.md
|-- 02-exemplar-analysis-<timestamp>.md
|-- 02-title-options-<timestamp>.md
|-- 03-venue-positioning.md
|-- venue-profiles/
|   `-- <custom-venue>.md
|-- 04-simulated-review.md
|-- 05-final-manuscript-<timestamp>.md
|-- 05-final-manuscript-<timestamp>.tex
|-- 05-final-manuscript-<timestamp>.pdf
|-- 06-response-letter.md
|-- 07-pre-submission-check.md
`-- 99-summary-report.md

Write order:

1. Generate or inspect the result.
2. Save the corresponding file.
3. Update .workspace.yml.
4. Report what was saved, what remains risky, and what decision is needed next.

Main Modes

Mode	Typical trigger	Output
Research positioning	"find the gap", "position this contribution"	01-research-positioning.md
Writing assistance	"draft this paragraph", "make this more academic"	text response or writing audit file
Writing-quality audit	"audit this draft", "check writing quality"	02-writing-audit.md
Venue positioning	"target ACL", "fit this for NeurIPS"	03-venue-positioning.md
Simulated review	"simulate peer review", "would this be rejected?"	04-simulated-review.md
Manuscript assembly	"assemble the manuscript", "compile the paper"	05-final-manuscript-<timestamp>.*
Reviewer response	"write a response letter", "handle these reviews"	06-response-letter.md
Pre-submission check	"pre-submission check", "ready to submit?"	07-pre-submission-check.md
Asynchronous full pass	"run the full pass", "audit everything"	99-summary-report.md

Submodes:

Submode	Parent mode	Typical trigger	Output
Exemplar analysis	Writing-quality audit	"analyze this great paper"	02-exemplar-analysis-<timestamp>.md
Title engine	Writing assistance	"give title options"	02-title-options-<timestamp>.md
Paragraph audit	Writing assistance	"audit this paragraph"	02-writing-audit-<timestamp>.md or direct reply

Before executing a mode, read the corresponding section in references/workflows.md.

Core Boundaries

• Do not invent research content, citations, experiments, datasets, statistics, or reviewer comments.
• Do not treat fluency as evidence. Claims must remain traceable to source material.
• Do not claim venue compliance unless current official requirements or user-provided author instructions have been checked.
• Research positioning must be based on searched or provided literature evidence; if the gap is weak, say so.
• quick_scan.py is a mechanical pre-scan, not a complete writing-quality audit.
• Simulated review requires independent reviewer passes when the platform supports subagents. If true isolation is unavailable, disclose the downgrade.
• Asynchronous full pass never replaces the user's final research, submission, authorship, or ethics decisions.

Done Criteria

A mode is done only when:

• the workspace state is clear;
• required materials and missing assumptions are recorded;
• generated claims remain anchored to user-provided or searched evidence;
• output has been saved when the mode promises a file;
• remaining risks and next decisions are explicit.