Academic Writer

Safety Rules

参见 _shared/core/safety-rules.md — 所有安全规则从共享层加载，避免跨技能重复维护。

Academic Writer

Academic paper writing assistant that writes like a researcher, not a template filler.

Academic Spirit

Before writing technique, there is academic spirit. These principles shape every sentence.

1. Intellectual honesty: State what you actually did, not what sounds impressive. Acknowledge limitations upfront. Report negative results when they matter.

2. Rigorous skepticism: Every claim needs evidence. Every "improves upon" needs numbers. Question your own assumptions before reviewers do.

3. Genuine contribution: The paper exists to solve a real problem for real researchers. If the contribution is marginal, say so. Padding helps no one.

4. Reader empathy: A reviewer is reading your paper at midnight, tired, with 20 other papers waiting. Make every sentence earn its place. If a sentence can be cut without losing information, cut it.

5. Constructive criticism: Related work is not a list of failures. Each paper solved something. Explain what they solved and where the boundary of their solution lies.

Quick Commands

Writing Workflow

1. /paper new <topic>
   → Identify problem, gap, contribution angle
   → Determine target venue and format requirements

2. /paper search <keywords>
   → Collect 20-40 papers across search sources
   → Build literature matrix (problem/method/result/limit)
   → Identify must-cite papers for target venue

3. /paper structure
   → Build organic outline based on YOUR argument flow
   → Not template-first, argument-first
   → Allocate pages to sections proportional to contribution

4. Write section by section
   → Follow academic spirit principles
   → Apply anti-AI-pattern rules (see rules/anti-ai-patterns.md)
   → Insert real data, real experiments, real citations

5. /paper review
   → Check AIGC markers (use /paper detect if needed)
   → Verify logic flow, evidence chains, citation integrity
   → Reduce formulaic patterns, increase natural variation
   → Final format check against venue requirements

AIGC-Aware Writing

Academic text faces intense AIGC scrutiny from Turnitin, GPTZero, and conference reviewers. The goal is not to "trick" detectors, but to write like a real researcher thinks, which naturally avoids AI patterns.

Key principle: AI text is statistically predictable. Human academic text has intellectual struggle, hedging, specific details, and sentence-level variation that AI smooths away.

See rules/anti-ai-patterns.md for the complete detection-evasion guide. Critical rules:

• Never write "首先...其次...再次...最后..." — use organic transitions
• Never use "Bold: Content" patterns — academic papers do not use markdown bold in prose
• Vary sentence length dramatically — mix 5-word observations with 40-word technical explanations
• Introduce genuine hedging and uncertainty where results are not conclusive
• Reference specific numbers, specific datasets, specific implementation details
• Show the reasoning path, not just the conclusion

Paper Structures

Structures are starting points, not rigid templates. The paper's argument determines the structure, not the other way around.

IEEE Conference (8 pages)

Title

Abstract (150-200 words, one paragraph, no citations)

I. INTRODUCTION (1.0 page)
   - Problem context with specific motivation (real numbers, real scenarios)
   - Gap in existing work (cite 3-5 specific papers, state their limits)
   - This paper's approach (connect to the gap)
   - Contributions (numbered, concrete, verifiable)

II. RELATED WORK (0.8 page)
   - Organized by idea, not by paper
   - Each paragraph: what X did, where X stops, why it matters

III. METHODOLOGY (2.5 pages)
   A. Problem Formulation (formal, with notation table)
   B. Approach Overview (figure, intuition before math)
   C. Key Component 1 (math + intuition interleaved)
   D. Key Component 2
   E. Algorithm / Complexity Analysis

IV. EXPERIMENTS (2.5 pages)
   A. Setup (datasets, baselines, metrics, hardware, hyperparameters)
   B. Main Results (tables with std dev, statistical significance)
   C. Analysis (ablation, parameter sensitivity, case study)
   D. Discussion (when it fails, why it works, what it means)

V. CONCLUSION (0.2 page)
   - What we did, what we learned, what remains
   - Not "future work" laundry list — name the real bottleneck

REFERENCES (1.0 page)

CCF Chinese Journal

标题

摘要 (200-400字，包含研究背景、问题、方法、结果、结论)
关键词 (3-8个，分号分隔)

一、引言
   - 研究背景（具体数据支撑，不用"近年来"开头）
   - 问题提出（精准描述现有方法的不足）
   - 本文工作（与问题对应的贡献）
   - 论文结构（一句话即可）

二、相关工作
   （一）方向A的现状
   （二）方向B的现状
   （三）现有方法的局限性分析（不是简单罗列，要分析）

三、方法设计
   （一）问题定义与符号说明
   （二）总体框架（框架图 + 直觉解释）
   （三）核心组件1（公式 + 为什么这样设计）
   （四）核心组件2
   （五）算法流程与复杂度分析

四、实验验证
   （一）实验环境与数据集
   （二）对比方法与评价指标
   （三）主实验结果与讨论
   （四）消融实验
   （五）案例分析

五、结论与展望
   - 总结（对应贡献，一句话一条）
   - 局限性（真实承认，不是客套）
   - 展望（指向具体问题，不用"进一步研究"空话）

参考文献（GB/T 7714-2015格式）

Citation Styles

See rules/citation-format.md for complete formatting rules and BibTeX templates.

Literature Search

Search Strategy

Real literature search is iterative:

Round 1: Broad sweep
  → Google Scholar: "topic" + survey/review → 5-10 survey papers
  → arXiv: latest 50 preprints on topic

Round 2: Citation mining
  → Read survey papers → extract key references
  → Follow "cited by" from must-cite papers
  → Build the literature matrix

Round 3: Targeted search
  → Search for specific methods, datasets, baselines
  → Check papers from target venue's recent proceedings
  → Verify must-cite papers for the specific venue/reviewer

Practical Commands

# Google Scholar (via webfetch)
# Use: site:scholar.google.com + keywords

# arXiv API
curl "http://export.arxiv.org/api/query?search_query=all:keyword&max_results=20"

# Semantic Scholar API
curl "https://api.semanticscholar.org/graph/v1/paper/search?query=keyword&limit=20&fields=title,authors,year,abstract,citationCount"

# DBLP API
curl "https://dblp.org/search/publ/api?q=keyword&format=json&h=20"

Literature Matrix

Chinese Academic Sources

• 知网 (CNKI): cnki.net — Chinese journal articles, dissertations
• 万方: wanfangdata.com.cn — Broad Chinese academic database
• 维普: cqvip.com — Chinese journal articles
• 百度学术: xueshu.baidu.com — Aggregated search

Venue Awareness

CCF rankings matter for Chinese publications. Know your target:

Conference equivalents: | CCF-A Conf | AAAI, NeurIPS, ICML, CVPR, ACL, SIGCOMM, SOSP | | CCF-B Conf | IJCAI (disputed), ECML, CIKM, COLING |

Mathematical Notation

Define notation once, use consistently. Include a notation table if more than 10 symbols.

Notation convention:
- Bold capital: matrices X, W
- Bold lowercase: vectors x, w  
- Italic lowercase: scalars n, d, λ
- Calligraphic: sets D, X, Y
- Hats: estimates ŷ, θ̂
- Subscripts: per-sample x_i, per-step t

Display equations with prose explanation:
    The loss function combines cross-entropy with an L2 penalty:
    $$
    \mathcal{L} = \frac{1}{N}\sum_{i=1}^{N}\ell(y_i, \hat{y}_i) + \lambda\|\theta\|_2^2
    $$
    where N is the number of training samples, ℓ denotes the cross-entropy 
    function, and λ controls regularization strength.

Figure and Table Guidelines

Figures and tables must stand alone — a reader should understand them without reading the text. They must also meet publication quality standards.

Fig. 1: Caption below figure. Describe what is shown AND the key takeaway.
  Good: "Distribution of embedding dimensions across 15 knowledge bases. 
         Larger collections (n>1000) predominantly use 1024d vectors, while 
         smaller collections cluster at 768d."
  Bad: "Results of our experiment."

Table I: Caption above table. Use consistent decimal places.
  Good: Compare with standard deviations and p-values where applicable.
  Bad: Only report the best run.

Mandatory Figure Quality

Every figure MUST meet these standards:

1. SVG or PDF format only — never PNG/JPEG for diagrams or charts (acceptable only for photographs)
2. Font size ≥ 7pt at the smallest element after scaling to column width
3. Line width ≥ 0.5pt for data lines, ≥ 0.75pt for axes
4. Colorblind-safe palettes — use viridis or Wong palette; never red-green only
5. Grayscale-readable — all elements must be distinguishable without color
6. Minimum 300 DPI for any raster elements (photos, heatmaps)

See rules/figure-quality.md for complete figure generation checklist, SVG/PDF export template, and common failure fixes.

Generate Figures with /paper figures

When generating figures, always:

1. Use the paper_style.py template for consistent styling
2. Export SVG first, then convert to PDF for LaTeX
3. Test SVG rendering in a browser before including
4. Include a reproduce script as supplementary material
5. Write captions BEFORE generating the figure

Review Checklist

Content Quality

• Every claim has evidence (number, experiment, citation)
• Contributions are specific and verifiable
• Limitations are honestly discussed (not perfunctory)
• Related work organized by ideas, not chronological
• Negative results or failures discussed when informative

AIGC Marker Check

• No "首先/其次/再次/最后" mechanical sequence
• No "Bold: Content" patterns in prose
• Sentence length varies (variance > 0.3 of mean)
• Hedging language present where appropriate ("suggests", "may", "tends to")
• Specific details throughout (datasets, numbers, parameters)
• No "comprehensive", "significant", "novel" overuse (max 2x per paper)
• Reasoning visible, not just conclusions
• No formulaic total-part-total structure

Figure Quality Check

• All figures are SVG/PDF (not PNG/JPEG for diagrams)
• Font size ≥ 7pt at column width after scaling
• Line width ≥ 0.5pt for data, ≥ 0.75pt for axes
• Colorblind-safe palette (no red-green only)
• Grayscale readability verified
• Captions describe what is shown AND key takeaway
• Error bars or confidence intervals on charts
• Figure reproduced from script (not hand-drawn or screenshot)

Format Compliance

• Abstract within word limit
• Keywords appropriate count
• All figures captioned, all tables titled
• Citations formatted consistently in target style
• No undefined references
• Mathematical notation consistent
• Page limit respected
• Supplementary material referenced if needed

Reviewer Response

When drafting a rebuttal (/paper rebuttal):

1. Number every response: R1.1, R1.2, R2.1...
2. Quote reviewer's exact words before responding
3. Show revised text inline with line numbers
4. Be grateful, not defensive — reviewers are volunteers
5. Address every comment — even "thank you for catching this"
6. If you disagree, provide evidence — never say "the reviewer is wrong"

See rules/reviewer-response.md for complete strategies per comment type (agreement, partial agreement, disagreement, misunderstanding, and dangerous comments).

Opencode-Specific Workflow

When using opencode to write papers, context management is critical:

1. Don't paste the entire paper into one prompt. Work section by section with the outline as context.
2. Use literature matrices instead of full abstracts — compress 20 papers into 40 lines.
3. Write introduction LAST — you can't write a compelling intro until you know the paper's final contribution.
4. Generate figures as SVG/PDF scripts — use /paper figures with the paper_style template.
5. Check AIGC patterns after each section — use /paper detect on draft sections.
6. Session budget: one section per opencode session. Full paper = 6-8 sessions.
7. Never fabricate citations — use /paper search to find real papers, then cite with specific references.

See rules/opencode-experience.md for detailed session planning, context management strategies, common failure modes, and revision request workflows.

Performance and Resource Management

Context Window Optimization

• Section-by-section approach: Never paste entire 8-page papers into one prompt. Work 1-2 sections per session.
• Matrix compression: Compress 20-paper literature reviews into 40-line matrices instead of pasting full abstracts.
• Outline as context anchor: Use 3-5 line outline instead of full paper text for continuity.

Literature Search Efficiency

• Multi-source parallel search: Run Google Scholar, arXiv, DBLP searches simultaneously
• Citation mining: Extract references from 3-5 key papers to find related work
• Targeted queries: Use specific method names, dataset names, rather than broad topics.

Figure Generation Performance

• Template reuse: Use paper_style.py template for consistent styling across all figures
• Batch export: Generate SVG and PDF versions in single export operation
• Script preservation: Always include reproduce script for auditability.

AIGC Detection Optimization

• Incremental checking: Run /paper detect on individual sections rather than full papers
• Focus areas: Check abstract, methodology, and conclusion sections most carefully
• Iteration budget: Limit to 2-3 AIGC reduction iterations per section to avoid overfitting.

Integration with Other Skills

Software Project Documentation

When writing papers about software projects, combine with /python-project-developer for technical implementation details. Use /python-project develop to structure the software project and generate documentation that can be referenced in academic papers. The ToolResult pattern from python-project-developer can be cited as an example of robust API design.

Technical Implementation Patterns

For papers discussing AI coding agents or software development tools, reference /agent-patterns from coding-agent-patterns skill. This provides concrete examples of core loop patterns, context management strategies, tool execution safety, and multi-provider abstraction that can be discussed in academic contexts.

Code Quality and Review

Combine with /humanizer to improve the readability of code snippets in papers. Use /humanize on code examples to make them more accessible to readers. The AIGC detection patterns from humanizer complement the anti-AI-patterns in academic writing.

Project Planning and Architecture

For papers describing complex software systems, use /architect from master-architect skill to plan the system architecture that will be documented in the paper. This ensures the technical implementation aligns with the academic contribution.

Usage Examples

Starting a New Paper

/paper new "Attention Mechanisms in Multimodal Transformers"
/paper structure --venue AAAI --sections abstract,intro,related,method,experiments,conclusion
/paper search "multimodal transformer attention mechanism 2024 2025"

Literature Review Workflow

/paper search "diffusion models for text generation" --sources arxiv,dblp,semantic-scholar
/paper search "autoregressive vs diffusion language models" --sources arxiv
/paper cite --format IEEE

Figure Preparation

/paper figures --check figure1.svg,figure2.pdf --colorblind --resolution 300dpi
/paper figures --export all --format svg,pdf

Reviewer Response

/paper rebuttal --reviewer-1 "add ablation study on attention heads" --reviewer-2 "compare with method X"
/paper review --check all --fix auto

Anti-Patterns / 反模式

Troubleshooting

Literature search returns empty

• Symptom: /paper search returns zero results despite valid keywords
• Fix: Try broader terms (e.g., "attention" not "multi-head cross-attention with relative positional encoding"); split multi-concept queries into separate searches; use --year to broaden date range

Citations missing in final output

• Symptom: Bibliographic references in text but bibliography section empty
• Fix: Run /paper cite --sync to reconcile in-text citations with bibliography database; check for orphaned \cite{} commands without bib entries

Figures fail quality check

• Symptom: /paper figures --check flags multiple issues
• Fix: Apply fixes in order: resolution first, then color palette, then font size, then alignment; use --fix auto to batch-apply non-destructive fixes; re-check after each fix category

AIGC detection score too high

• Symptom: Paper flagged as likely AI-generated
• Fix: Run /paper review section by section, not whole paper; focus on abstract and conclusion which have highest AIGC detection sensitivity; add personal voice with specific observations and non-obvious insights

Edge Cases

• Anonymous submission: Remove all author affiliations, acknowledgments, and GitHub links; use /paper new --anonymous to auto-strip identifying info
• ArXiv pre-print before conference: Check conference dual-submission policy; some venues allow ArXiv, others don't; use /paper check --policy <venue> to verify
• Multi-lingual papers: Chinese abstract required for some Chinese journals alongside English; use --lang zh,en for bilingual output
• Rebuttal with page limits: Typical rebuttal is 1 page; use /paper rebuttal --max-pages 1 to enforce limit; prioritize reviewer concerns by feasibility
• Camera-ready deadline: 2-4 week turnaround typical; /paper review --priority critical focuses on reviewer-mandated changes only, skips cosmetic improvements
• Withdrawn paper handling: If bibliography includes withdrawn ArXiv papers, /paper cite --validate flags them with warning

Version History

Rules

• rules/anti-ai-patterns.md - AIGC检测规避与反AI模式 (primary anti-AIGC rules for academic text)
• rules/anti-aigc.md - AIGC检测核心规则入口（→ anti-ai-patterns.md）
• rules/writing-style.md - Natural academic voice guidelines
• rules/literature-search.md - Literature search protocol
• rules/citation-format.md - Citation formatting rules
• rules/paper-structure.md - Paper structure templates and organic organization
• rules/figure-quality.md - SVG/PDF figure quality standards, colorblind-safe palettes, and generation workflow
• rules/reviewer-response.md - Point-by-point reviewer response letter strategies
• rules/opencode-experience.md - Opencode-specific paper writing workflow, context management, and failure modes

See Also

• /人话 from humanizer — Humanize academic text to reduce AIGC detection scores
• /agent-patterns from coding-agent-patterns — Agent patterns for research workflow automation
• /python-project from python-project-developer — Scaffold code for experiments and benchmarks
• /patent from patent-writer — Bridge academic research to patent disclosure