---
id: gh-research-writing-harness
name: "research-writing-harness"
url: https://skills.yangsir.net/skill/gh-research-writing-harness
author: 4everWZ
domain: science
tags: ["research-writing", "academic-literature", "citation-management", "paper-generation", "evidence-management"]
install_count: 520
rating: 4.20 (120 reviews)
github: https://github.com/4everWZ/research-writing-harness
---

# research-writing-harness

> 这是一个专为深度学习研究项目设计的按需论文写作与证据管理工具。它协助用户进行文献搜索、引用管理、论文索引维护、研究思路提炼、代码转论文章节、引用审计以及论文交付准备。它不是一个自主研究系统，仅在用户明确请求时激活，旨在提升学术写作的效率和质量。

**Stats**: 520 installs · 4.2/5 (120 reviews)

## Before / After 对比

### 加速学术论文写作与管理

**Before**:

研究人员在手动管理学术文献、引用、撰写论文草稿时，耗费大量时间且易出错，导致代码、论点和参考文献之间存在不一致。大量精力被行政性工作而非核心研究占据。

**After**:

通过自动化和简化文献管理、引用审计以及将研究成果（代码、实验）转化为结构化论文章节，确保了内容的一致性并显著减少了手动工作量，使研究人员能更专注于科学贡献。

| Metric | Before | After | Change |
|---|---|---|---|
| 文献管理与论文初稿时间 | 60小时 | 10小时 | -83% |

## Readme

# Research Writing Harness

## Purpose

Use this skill as an on-demand paper-writing and evidence-management layer for an existing or evolving deep learning research project.

It is not an autonomous research system. It must not run as a default research layer. It should activate only when the user explicitly requests literature search, citation management, idea refinement from papers, repo-to-paper writing, citation audit, or paper handoff.

The main implementation harness owns code, training, debugging, validation, experiment execution, and code handoff. This skill owns literature evidence, paper index maintenance, claim tracking, Markdown paper sections, citation audit, and paper-writing handoff.

## Non-Goals

Do not use this skill to:

- autonomously conduct research;
- automatically run experiments;
- automatically write paper prose after literature search;
- invent results, ablations, metrics, captions, or conclusions;
- claim novelty, superiority, or SOTA without explicit evidence;
- change datasets, splits, metrics, protocols, baselines, or contribution framing without user confirmation;
- optimize for novelty over evidence quality, reproducibility, and real effect.

## Required Flat Workspace

For each paper-like project, use one flat workspace under `docs/`:

```text
docs/<paper_slug>/
  README.md
  venue_profile.md
  paper_index.md
  references.bib
  claims.md
  idea_log.md
  intro.md
  related_work.md
  method.md
  experiments.md
  results_tables.md
  limitations.md
  figures.md
  handoff.md
  papers/
  notes/
```

After the target venue/outlet is confirmed, the workspace folder may be renamed with a double-underscore suffix, for example:

```text
docs/fire_mamba_ir__cvpr/
docs/hwdown_deim__tip/
docs/vla_agent_eval__corl/
```

Use `venue_profile.md` to record venue/outlet assumptions. Do not create many venue-specific writing templates.

## Core Files

Treat these files as the central state:

- `paper_index.md`: indexed literature and evidence quality;
- `claims.md`: claim ledger linking draft claims to literature, code, experiments, user decisions, or explicit assumptions;
- `idea_log.md`: literature-driven idea refinement and rejected options;
- `venue_profile.md`: confirmed or provisional target venue/outlet and writing tendencies.

Do not write unsupported academic claims into paper sections without recording them in `claims.md`.

## Progressive Loading Policy

Load the smallest useful reference file. Do not read every reference by default.

### Load 0: Router only

Use only this `SKILL.md` when deciding whether the skill should apply, selecting a task route, or initializing a workspace.

### Load 1: Single task reference

Use one reference file when the task is narrow:

- Literature collection or idea refinement: load `references/literature.md`.
- Repo/code/config to paper section: load `references/repo-to-paper.md`.
- Citation/claim checking: load `references/citation-audit.md`.
- Paper handoff: load `references/handoff.md`.

### Load 2: Task reference plus templates

Load the relevant template from `assets/templates/` only when writing or validating that file.

Examples:

- `paper_index.md` work: load `references/literature.md` plus `assets/templates/paper_index.md`.
- Method section work: load `references/repo-to-paper.md` plus `assets/templates/method.md`.
- Claim audit: load `references/citation-audit.md` plus `assets/templates/claims.md`.

### Load 3: Cross-file consistency

Use Load 3 when the user asks for consistency across literature, claims, and sections. Inspect only the necessary workspace files, normally `paper_index.md`, `claims.md`, the target section, and `venue_profile.md`.

### Load 4: Major paper-state update

Use Load 4 only for Tier A/B changes or explicit paper handoff. Inspect the relevant workspace state and update `handoff.md` if required.

## Task Routing

### Literature collection

Use `references/literature.md` when the user asks to search, collect, index, classify, download, or summarize academic papers, build/update `paper_index.md`, add BibTeX, or create reading notes.

Do not write introduction, related work, or method prose after literature collection unless the user explicitly asks for section writing.

### Idea refinement from literature

Use `references/literature.md` when the user asks to refine an idea using papers.

Write candidate refinements to `idea_log.md`. Do not modify code. Do not convert suggestions into paper claims without evidence and user decision.

### Repo-to-paper writing

Use `references/repo-to-paper.md` when the user asks to convert implemented code, configs, architecture notes, or experiment setup into Markdown paper sections.

Write only the requested section. Update `claims.md` for nontrivial claims introduced by the section.

### Citation audit

Use `references/citation-audit.md` when the user asks whether citations support claims, whether related work is distorted, whether novelty is overclaimed, or whether claims are unsupported.

### Handoff

Use `references/handoff.md` when the user asks for paper handoff or when a Tier A/B change has occurred.

## Source Hierarchy Summary

Use the full policy in `references/literature.md`. The short version is:

1. Prefer recent formal peer-reviewed papers from the last 1-3 years in top venues and top journals recognized by the relevant subfield.
2. Include older foundational work, strong baselines, community-standard methods, datasets, metrics, and protocols when still relevant.
3. Use recent arXiv only as frontier supplement, not as sole theoretical support or sole evidence for key conclusions.
4. Strictly downgrade or exclude MDPI, Hindawi, Frontiers, isolated low-quality preprints, marketing-like sources, weakly reproducible works, and sources with unclear academic consensus.
5. Venue rank is not evidence. Final evidence strength depends on relevance, method clarity, baseline strength, evaluation fairness, dataset/metric/split transparency, code/reproducibility support, and claim-evidence alignment.

## Confirmation Checkpoints

Ask the user before:

- changing contribution framing;
- declaring novelty or SOTA;
- choosing the final method direction;
- removing or replacing strong baselines;
- changing dataset, split, metric, or evaluation protocol;
- converting speculative refinements into final paper claims;
- downgrading or excluding a source if that decision materially affects the paper narrative.

## Results Policy

Do not write experimental results unless the user provides the numbers.

For results, you may only:

- create table structures;
- define metric columns;
- prepare neutral captions;
- mark missing cells as `TODO`;
- list required experiments.

## Handoff Policy

Do not update `handoff.md` for every small edit.

Update `handoff.md` for:

- Tier A: paper direction, contribution framing, method direction, baseline strategy, or evaluation protocol changes;
- Tier B: completed literature collection, major section draft, important idea refinement, strong similar-work discovery, or major evidence gap discovery.

Do not update `handoff.md` for Tier C local edits such as wording, formatting, single BibTeX fixes, or one citation addition.

## Scripts

Use scripts only for low-risk local structure checks:

```bash
python scripts/init_paper_workspace.py docs/<paper_slug>
python scripts/validate_workspace.py docs/<paper_slug>
python scripts/validate_paper_index.py docs/<paper_slug>/paper_index.md
python scripts/quick_validate_skill.py .
```

Do not add or rely on scripts that scrape Google Scholar, automatically summarize PDFs, or automatically write related work.


---

# research-writing-harness

A Codex/OpenAI-style global skill for on-demand literature-grounded paper writing and evidence management.

This skill is not an autonomous research system. It is a paper-writing evidence layer for existing or evolving research projects.

## Install

Place the `research-writing-harness/` folder in your Codex skills directory, for example:

```text
$CODEX_HOME/skills/research-writing-harness/
```

Restart Codex after installing or updating the skill.

## Standard Skill Shape

```text
research-writing-harness/
  SKILL.md
  agents/openai.yaml
  references/
  assets/templates/
  scripts/
```

`SKILL.md` is at the skill folder root and has YAML frontmatter with only `name` and `description`.

## What It Does

- Search and curate literature.
- Maintain `paper_index.md`.
- Maintain `references.bib`.
- Maintain `claims.md`.
- Refine ideas using literature without changing code automatically.
- Convert implemented code/config/design into Markdown paper sections.
- Audit claims and citations.
- Prepare Tier A/B paper handoff.

## What It Does Not Do

- Does not auto-run research.
- Does not always activate.
- Does not write paper sections after literature search unless explicitly requested.
- Does not invent results.
- Does not claim novelty/SOTA without evidence.
- Does not use venue rank alone as evidence quality.

## Initialize a Workspace

```bash
python scripts/init_paper_workspace.py docs/fire_mamba_ir
```

After target venue/outlet is confirmed:

```bash
python scripts/init_paper_workspace.py docs/fire_mamba_ir --venue CVPR --suffix-venue
```

This creates:

```text
docs/fire_mamba_ir__cvpr/
  README.md
  venue_profile.md
  paper_index.md
  references.bib
  claims.md
  idea_log.md
  intro.md
  related_work.md
  method.md
  experiments.md
  results_tables.md
  limitations.md
  figures.md
  handoff.md
  papers/
  notes/
```

## Validate

```bash
python scripts/quick_validate_skill.py .
python scripts/validate_workspace.py docs/fire_mamba_ir
python scripts/validate_paper_index.py docs/fire_mamba_ir/paper_index.md
```


---
*Source: https://skills.yangsir.net/skill/gh-research-writing-harness*
*Markdown mirror: https://skills.yangsir.net/api/skill/gh-research-writing-harness/markdown*