academic-writing-cs（学术写作 cs）

Academic Writing for Computer Science

Overview

This skill provides end-to-end support for writing high-quality computer science research papers. It focuses on constructing clear, compelling technical narratives while adhering to field-specific conventions.

Core Philosophy:

• Academic papers are narrative arcs (Problem → Solution → Evidence → Implications), not template fill-ins
• Clarity comes from structure: place familiar information first, new information last
• Every design choice must be justified; every claim must be supported

Scope:

• Conference papers (6-12 pages, competitive venues)
• Journal articles (15-30 pages, comprehensive)
• Thesis chapters (flexible length, deep coverage)
• All CS subfields: AI/ML, Systems, Theory, HCI, Security, etc.

When to Use This Skill

Invoke this skill when:

• Planning paper structure and narrative flow
• Drafting any section (Abstract, Introduction, Methods, Results, Discussion, Conclusion)
• Revising for clarity, coherence, or compliance with venue requirements
• Reviewing sentence-level writing for clarity issues
• Seeking CS-specific conventions (notation, figures, citations)
• Checking completeness with section-by-section quality checklists
• Responding to reviewer comments

Workflow Decision Tree

Stage 1: Planning and Structure

When starting a new paper or major revision:

1. Define the Narrative Arc Reference: references/narrative_framework.md — Read the "Core Principle" and "Section-Level Narrative Structure" sections to understand how to structure the paper's story.

- What problem does this solve, and why does it matter? (1-2 sentences) - What is the single main contribution? (1 sentence) - What are the 3 key results that support the contribution? - What are the main limitations?

1. Identify Target Venue and Constraints Reference: references/cs_conventions.md (Section 8: Venue-Specific Guidelines, Section 5: Subfield-Specific Conventions)

- Conference or journal? - Page limits, formatting requirements, anonymization rules? - Subfield conventions (ML vs. Systems vs. Theory)?

1. Outline Section-by-Section Tool: Use assets/section_checklists.md (Quick Pre-Draft Planning Checklist) to ensure all key questions are answered before writing begins.

- For each major section, define: - What is the purpose of this section? - What are the 2-3 key points to convey? - What figures/tables will support this?

Stage 2: Drafting

For each section, follow this process:

Abstract

1. Use the 4-sentence structure: Context → Gap → Contribution → Impact
2. Check against assets/section_checklists.md (Abstract Checklist)
3. Ensure it's self-contained and within word limit (150-250 words)

Common mistakes:

• Vague contribution: "We improve X" → Be specific: "We achieve 15% higher accuracy"
• No concrete results: Always include numbers/metrics

Introduction

1. Follow the funnel structure: Broad → Narrow → Specific

- Para 1: Problem domain and importance - Para 2-3: Specific problem, motivation, why existing work falls short - Para 4: Gap statement ("However, existing approaches lack...") - Para 5: Contribution overview (what this paper provides) - Para 6: Results summary (2-3 concrete findings) - Para 7: Paper organization (optional)

1. Key requirement: By the end of paragraph 4-5, the reader must clearly understand the contribution.
2. Include at least one figure (architecture or key result) for ML/systems papers.
3. Check against assets/section_checklists.md (Introduction Checklist)

Reference: references/narrative_framework.md (Introduction section) for detailed guidance and examples.

Related Work

1. Organize thematically (not chronologically): Group into 3-5 categories
2. For each category:

- Describe the general approach - Cite 3-5 representative works with 1-sentence descriptions - Point out limitations relevant to your contribution

1. End with positioning paragraph: "In contrast to [X], our approach..."

- Clearly articulate differences and advantages

1. Check against assets/section_checklists.md (Related Work Checklist)

Common mistakes:

• Laundry list of citations without synthesis
• Failing to position your work relative to prior work
• Being dismissive (respect prior work while differentiating)

Methodology

1. Dual objectives:

- Reproducibility: Enough detail for reimplementation - Intuition: Explain why the approach works

1. Structure varies by paper type:

- ML/AI papers: Problem Formulation → Overview + Figure → Detailed Design → Implementation → Complexity - Systems papers: Architecture Overview → Component Design → Key Mechanisms → Implementation - Theory papers: Formal Definitions → Main Results (theorems) → Proof Sketch

1. Always include:

- Clear notation (define all symbols on first use) - High-level overview before diving into details - Justification for design choices (or defer to Ablations)

1. Check against assets/section_checklists.md (Methodology Checklist)

Reference: references/narrative_framework.md (Methodology section) and references/cs_conventions.md (Section 1: Notation and Mathematical Writing)

Experiments/Results

1. Experimental Setup (subsection):

- Datasets: Size, splits, preprocessing - Baselines: What you compare against (with citations) - Metrics: What you measure and why - Hardware/Software: Infrastructure and versions - Hyperparameters: How selected

1. Main Results (subsection):

- Table/figure showing primary comparison - Text: "Table 1 shows that our method outperforms..." - Highlight key findings with concrete numbers - Report statistical significance (confidence intervals, p-values, or std dev)

1. Ablation Studies (subsection, critical):

- Demonstrate necessity of each component - Table: effect of removing/modifying components

1. Analysis (subsection):

- Where does the method excel? Where does it fail? - Qualitative analysis, error analysis, failure cases

1. Computational Cost (if relevant):

- Training time, inference time, memory usage - Comparison with baselines

1. Check against assets/section_checklists.md (Experiments/Results Checklist)

Reference: references/narrative_framework.md (Experiments/Results section)

Discussion

1. Summarize findings (1 para): Restate key results
2. Interpret results (1-2 paras): Why does the method work? What insights?
3. Acknowledge limitations (0.5-1 para): Be honest about scope and failure cases
4. Broader implications (0.5-1 para): Impact on the field, applications, future directions
5. Check against assets/section_checklists.md (Discussion Checklist)

Tone: Balanced—confident but not overselling. Limitations increase credibility.

Conclusion

1. Restate contribution (1 para): Recap problem, solution, key findings
2. Broader impact (0.5 para): Significance and applications
3. Future work (0.5 para): Open questions and extensions

- Phrase as opportunities: "An interesting direction is..." (not "In future work, we will...")

1. Check against assets/section_checklists.md (Conclusion Checklist)

Do NOT: Introduce new ideas, copy-paste Abstract, or be vague.

Stage 3: Revision for Clarity

After drafting, apply sentence-level clarity principles:

The Three Golden Rules (Gopen & Swan)

1. Old Before New: Start sentences with familiar information; end with new information

- This creates coherent flow where each sentence builds on what came before

1. Subject-Verb Proximity: Keep the verb close to the subject

- Long gaps between subject and verb strain comprehension

1. Stress Position Power: Place the most important information at sentence end

- Readers remember and emphasize what comes at the end

Apply these rules systematically:

• For each paragraph, check that sentences flow (old-to-new)
• For each sentence, check that:

- Topic position (start) contains familiar info - Stress position (end) contains important new info - Verb appears soon after subject

Reference: references/sentence_clarity.md — Read this in full for detailed principles, examples, and common anti-patterns.

Practical Checklist:

• Familiar information at sentence start (topic position)
• Important new information at sentence end (stress position)
• Verb close to subject
• Active voice (unless passive is intentionally better)
• Parallel structures for parallel ideas

Common anti-patterns to fix:

• "Buried Verb" Syndrome: Converting verbs to nouns (nominalization)

- ❌ "The comparison of the methods is shown..." - ✅ "Table 1 compares the methods..."

• "Throat-Clearing": Weak starts like "It is important to note that..."

- ❌ "It is important to note that our method improves accuracy." - ✅ "Our method improves accuracy."

• "Dangling Emphasis": Ending sentences with weak elements

- ❌ "This approach significantly improves performance, as shown in [23]." - ✅ "As shown in [23], this approach significantly improves performance."

Stage 4: Polishing and Compliance

Language and Phrasing

When writing or revising specific academic functions, consult references/phrasebank.md:

• Introducing work: Establishing territory, identifying gaps, stating contributions
• Referring to sources: Integral vs. non-integral citations
• Describing methods: Sequential actions, conditional logic, implementation details
• Reporting results: Presenting findings, comparing baselines, interpreting
• Discussing findings: Explaining success, acknowledging limitations, stating implications
• Writing conclusions: Summarizing, broader impact, future work

General language functions:

• Being cautious (hedging): "may", "appears to", "likely"
• Being critical: Identifying weaknesses, questioning validity
• Compare and contrast: Similarity, difference
• Describing trends: Increasing, decreasing, stability
• Explaining causality: Causes, effects, conditions

Usage: Adapt templates to your context; don't copy verbatim. Vary expressions to maintain natural flow.

CS-Specific Conventions

Ensure compliance with field norms:

1. Notation:

- Define all symbols on first use - Use consistent conventions (bold for vectors, italic for scalars, etc.) - Integrate equations into sentences with punctuation

1. Figures and Tables:

- Reference all figures/tables in text before they appear - Self-contained captions - High-resolution, readable fonts (≥8pt) - Colorblind-friendly palettes

1. Citations:

- Follow venue citation style (author-year or numbered) - Cite all prior work you build on or compare against - Accurate and complete bibliography

1. Code and Reproducibility:

- State code availability - Provide sufficient implementation details - Report hyperparameters, random seeds, number of runs

1. Subfield-Specific Variations:

- ML/AI: Emphasis on ablations, statistical significance, computational cost - Systems: Architecture diagrams, throughput/latency, scalability - Theory: Formal definitions, theorems, proofs, complexity bounds - HCI: User studies, qualitative feedback, interface screenshots - Security: Threat models, attack scenarios, defense mechanisms

Reference: references/cs_conventions.md — Comprehensive guide covering notation, figures, citations, code, subfield norms, and venue requirements.

Quality Assurance

Before submission, use assets/section_checklists.md:

1. Section-by-Section Review:

- Run through each section's checklist - Ensure all required elements are present - Check for common pitfalls

1. Pre-Submission Checklist:

- Content completeness (all sections, figures, citations) - Formatting (venue template, page limits, margins) - Anonymization (if double-blind) - Reproducibility (sufficient detail, code availability) - Final quality checks (spell-check, grammar, co-author review)

1. Emergency Checklist (if deadline is imminent):

- Prioritize: Abstract, Introduction contribution statement, Main results table, At least one ablation, Readable figures, Correct bibliography

Stage 5: Responding to Reviews

After receiving reviewer feedback:

1. Analyze comments systematically:

- Categorize: Major issues (experiments, clarity, claims) vs. Minor issues (typos, formatting) - Prioritize: Address major issues first

1. Plan revisions:

- List all changes to be made - If experiments are requested, plan them carefully - If clarifications are needed, identify which sections to revise

1. Revise and respond:

- Address every comment (in rebuttal or revision) - Use respectful, professional tone - Clearly mark changes (if required by venue)

1. Check revised version:

- Ensure all changes are integrated - Re-run relevant checklists from assets/section_checklists.md (Revision Checklist) - Verify still within page limits

Reference: assets/section_checklists.md (Revision Checklist)

Key Resources Summary

Narrative and Structure

• references/narrative_framework.md: Core paper structure (Abstract, Introduction, Related Work, Methods, Results, Discussion, Conclusion). Use for understanding the narrative arc and section-specific guidance.

Sentence-Level Clarity

• references/sentence_clarity.md: Gopen & Swan principles (topic position, stress position, old-to-new flow). Use for revising individual sentences and paragraphs for maximum clarity.

Academic Phrases

• references/phrasebank.md: Templates for common academic writing functions (introducing work, citing sources, reporting results, discussing findings). Use when drafting or seeking variation in phrasing.

CS Conventions

• references/cs_conventions.md: Field-specific norms (notation, figures, citations, code, subfield variations, venue requirements). Use for ensuring compliance with CS writing standards.

Quality Checklists

• assets/section_checklists.md: Comprehensive checklists for every section, plus pre-submission, revision, and emergency checklists. Use for planning, reviewing, and final quality assurance.

Example Workflows

Workflow 1: Starting from Scratch

User: "I need to write a conference paper on my new semi-supervised learning method."

Process:

1. Planning (Stage 1):

- Define narrative arc: Problem (labeled data is expensive) → Solution (our semi-supervised method) → Evidence (experiments on 3 datasets) → Implications (reduces labeling cost) - Read references/narrative_framework.md (Core Principle) - Use assets/section_checklists.md (Quick Pre-Draft Planning Checklist)

1. Drafting (Stage 2):

- Abstract: 4-sentence structure (Context: deep learning needs data; Gap: labeling is expensive; Contribution: our method STCR; Impact: 82% accuracy with 10% labels) - Introduction: Funnel (broad: DL success → narrow: labeling cost → gap: existing semi-supervised methods lack X → contribution: STCR leverages consistency → results: 7% improvement) - Check each section against assets/section_checklists.md

1. Revision (Stage 3):

- Apply references/sentence_clarity.md principles to every paragraph - Ensure old-to-new flow, stress position usage

1. Polishing (Stage 4):

- Use references/phrasebank.md for varied phrasing - Ensure compliance with references/cs_conventions.md (ML/AI conventions) - Run Pre-Submission Checklist from assets/section_checklists.md

Workflow 2: Revising for Clarity

User: "My introduction is confusing. Reviewers said they couldn't understand the contribution."

Process:

1. Diagnose issue:

- Check against assets/section_checklists.md (Introduction Checklist) - Is the contribution stated clearly by paragraph 4-5? - Is the funnel structure followed (broad → narrow)?

1. Restructure if needed:

- Read references/narrative_framework.md (Introduction section) - Ensure: Opening → Background → Gap → Contribution → Results → Organization - Explicitly state: "In this paper, we present [X], which addresses [Y] by [Z]."

1. Revise at sentence level:

- Apply references/sentence_clarity.md principles - Check that each sentence flows from the previous one (old-to-new) - End key sentences with the important information (stress position)

Workflow 3: Drafting the Results Section

User: "How should I present my experimental results?"

Process:

1. Structure:

- Read references/narrative_framework.md (Experiments/Results section) - Follow: Setup → Main Results → Ablations → Analysis → Cost

1. Create tables/figures:

- Main results table: Methods (rows) vs. Metrics (columns) - Bold best results; include standard deviations - Check references/cs_conventions.md (Figures and Tables section)

1. Write accompanying text:

- "Table 1 shows that our method achieves X, outperforming the strongest baseline by Y%." - Use references/phrasebank.md (Section 4: Reporting Results) for phrasing

1. Quality check:

- Run through assets/section_checklists.md (Experiments/Results Checklist) - Ensure: Statistical significance, Ablations present, Analysis included

Workflow 4: Ensuring CS Compliance

User: "Is my notation and citation style correct for ICML?"

Process:

1. Check venue requirements:

- Read references/cs_conventions.md (Section 8: Venue-Specific Guidelines) - ICML uses numbered citations [1], double-blind review, LaTeX template

1. Notation:

- Read references/cs_conventions.md (Section 1: Notation and Mathematical Writing) - Ensure: Vectors are bold, scalars are italic, all symbols defined

1. Citations:

- Read references/cs_conventions.md (Section 3: Citations and References) - Use numbered format: "Method X [1] achieves..." - Anonymize self-citations for double-blind

1. Final check:

- assets/section_checklists.md (Pre-Submission Checklist → Compliance section)

Common Pitfalls and How to Avoid Them

Pitfall 1: Vague Contributions

Problem: "We improve performance on X." Solution: Be specific. "We achieve 15% higher accuracy than the strongest baseline on ImageNet."

Pitfall 2: Missing Ablations

Problem: Claiming design choices are important without evidence. Solution: Include ablation studies. Remove each component and measure the performance drop.

Pitfall 3: Poor Information Flow

Problem: Sentences feel disjointed; readers get lost. Solution: Apply old-to-new flow. Each sentence should start with information from the previous sentence. Reference: references/sentence_clarity.md

Pitfall 4: Weak Stress Position

Problem: Sentences end with citations or minor details. Example: ❌ "This approach significantly improves performance, as shown in [23]." Solution: ✅ "As shown in [23], this approach significantly improves performance."

Pitfall 5: Ignoring Limitations

Problem: Overselling without acknowledging scope or failure cases. Solution: Dedicate a paragraph in Discussion to honest limitations. This increases credibility.

Pitfall 6: Inconsistent Notation

Problem: Using x for input in one section, X in another. Solution: Define all notation upfront. Create a notation table (appendix) if needed. Reference: references/cs_conventions.md (Section 1)

Tips for Efficient Writing

1. Draft quickly, revise thoroughly:

- Don't aim for perfection in the first draft - Get ideas down, then refine structure and clarity

1. Write sections out of order:

- Start with Methods and Results (most concrete) - Then Introduction and Related Work - Finally Abstract and Conclusion

1. Use figures early:

- Create key figures (architecture, main results) before writing - Figures clarify your thinking and guide the narrative

1. Get feedback early:

- Share drafts with co-authors and colleagues - Mock reviews identify issues before submission

1. Iterate on structure:

- If a section feels wrong, revisit the narrative arc - Ensure every section advances Problem → Solution → Evidence → Implications

1. Use the checklists proactively:

- Before drafting a section, read the checklist to know what to include - After drafting, use the checklist to verify completeness

Advanced: Handling Special Cases

Writing for Top-Tier Venues

• Higher bar for novelty and rigor: Ensure the contribution is significant, not incremental
• Strong baselines: Compare against state-of-the-art, not just simple methods
• Comprehensive evaluation: Multiple datasets, extensive ablations, sensitivity analyses
• Polished presentation: High-quality figures, clear writing, consistent notation

Writing Rebuttals

• Address all concerns: Even if you disagree, engage respectfully
• Provide evidence: If reviewers doubt a claim, provide additional results or citations
• Be concise: Rebuttals have strict length limits; prioritize major issues
• Highlight changes: "We added an experiment (Table 3) showing..."

Writing Thesis Chapters

• More comprehensive: Deeper background, extended related work, lessons learned
• Narrative continuity: Ensure chapters connect (e.g., Chapter 3 builds on Chapter 2)
• Broader scope: Can include negative results and explorations that didn't pan out
• Use assets/section_checklists.md (Long-Form Paper Checklist)

Summary: The Golden Workflow

1. Plan the narrative: Problem → Solution → Evidence → Implications
2. Draft section-by-section: Use structure guidelines from references/narrative_framework.md
3. Revise for clarity: Apply principles from references/sentence_clarity.md
4. Polish and comply: Use references/phrasebank.md and references/cs_conventions.md
5. Quality check: Run through assets/section_checklists.md

Remember:

• Papers are stories, not templates
• Clarity comes from structure (old-to-new, topic/stress positions)
• Every claim needs evidence; every design choice needs justification
• Honest limitations increase credibility

When in doubt, ask:

• "Does this advance the narrative arc?"
• "Can a reader reproduce this?"
• "Is this claim supported?"
• "Is this the simplest, clearest way to express this?"

Getting Started

For a new paper:

1. Read references/narrative_framework.md (Core Principle)
2. Use assets/section_checklists.md (Quick Pre-Draft Planning Checklist)
3. Outline your paper's narrative arc in 4 sentences (Problem, Solution, Evidence, Implications)
4. Draft section-by-section, checking checklists as you go

For revising an existing draft:

1. Identify the issue (structure, clarity, compliance)
2. Consult the relevant reference file
3. Apply fixes systematically
4. Re-check with the appropriate checklist

For sentence-level issues:

1. Read references/sentence_clarity.md (Three Golden Rules)
2. Apply to each problematic paragraph
3. Check: Old-to-new flow, stress position usage, subject-verb proximity

Ready to write? Let's build a clear, compelling paper together.