| name | academic-paper-writing |
| description | Guides AI agents to write high-quality academic papers following top-tier conference standards, based on the "From Engineering Implementation to Scientific Argument" methodology.
Use this skill when:
- Writing or drafting academic papers for AI/systems conferences
- Designing paper structure, sections, or experimental methodology
- Structuring motivation, insight, and design narrative
- Creating figures, tables, or data visualizations for papers
- Performing literature review or managing citations
- Responding to reviewer comments or preparing rebuttal
- Any task involving academic writing workflow
Do NOT use for:
- Non-academic writing (blogs, marketing copy, documentation)
- Papers outside AI/systems domains without adaptation
- Simple formatting tasks that don't require domain expertise
|
Academic Paper Writing
From Engineering Implementation to Scientific Argument.
Core Philosophy
Writing = Building Arguments, Not Recording Work
Papers must construct an irrefutable argument system, not simply document what was done. The goal: make reviewers conclude "this problem is important, this method is sound, these results are credible."
Papers are judged by how hard they are to reject.
Common misconceptions:
- ❌ "I did a lot → should be recognized"
- ❌ "Many experiments → should be accepted"
- ❌ "Complex implementation → should be impressive"
The Argument Loop
Observation → Insight → Design → Evidence
Every section must form a verifiable takeaway (clear takeaway) that closes this loop.
Research vs Writing Logic
Thinking Pipeline (Chaotic Exploration)
Scenario → Problem → Insight → Idea → Design → Evaluation
- Exploratory: divergent, iterative, "trial"-driven
- Usually de-emphasizes insight
Writing Pipeline (Linear Storytelling)
Problem → Motivation(+Insight+idea-hint) → Design(+idea-expanded) → Evaluation
- Narrative: convergent, linear, "insight"-driven
- Tells a clean story, looks like natural derivation
- Extremely强化 insight
Writing requires logic reconstruction! Reduce entropy from "chaotic exploration" to "ordered narrative."
The Chain of Impact
Problem makes the paper meaningful
Insight makes the paper attractive
Idea makes the paper natural
Design makes the paper practical
Evaluation makes the paper undeniable
(Writing determines the outcome: same content, expression determines success)
Research discovers truth; writing makes truth appear inevitable.
Reviewer Psychology
Four Axioms
- Default Distrust: Reviewers assume most papers are poor quality
- Limited Patience: Key information must appear early; conclusions go first
- Emotional Judgment: Visual presentation directly affects acceptance decisions
- Logical Closure: Arguments must be self-consistent; gaps invite rejection
Reviewers seek flaws to REJECT. Give them reasons to ADVOCATE.
Key Insight
- Senior reviewers: won't over-focus on evaluation (insight/idea matter most)
- Young researchers will! ("Why can't I reproduce this effect?!")
Writing Workflow
Three-Phase Process
| Phase | Goal | Actions |
|---|
| Fast Output | Capture all ideas | Don't polish; write in Chinese then translate |
| Restructure | Rebuild logic | ~80% rewrite; refine figures; produce draft |
| Refine | Polish details | Adjust wording; add citations; fill pages; proofread |
Drafts are meant to be推翻; revision is the real creation. But the draft is the critical transition from "exploration" logic to "narrative" logic.
Role Transition
Early: Engineer (deep in details) → Late: Researcher (above details)
Engineer builds the system; Researcher builds the argument.
Hide the Engineer's pride, but keep the Engineer's rigor.
Mindset
- Don't overestimate your writing ability
- Don't pursue perfection (no paper is perfect)
- Self-satisfaction is baseline; peer-satisfaction is the goal
- Strive to: make misunderstanding impossible; make rejection difficult
- Writing is not a talent — it is a trained discipline
Paper Structure
Critical Sections (The Target)
These sections define the problem and build consensus:
Abstract: Mini Introduction; spend up to half on problem/motivation; show concrete numbers (improved by 80% not greatly improved)
Introduction: Not "introduction" but "strategic map" — make reviewers decide to read seriously
If intro fails, the paper fails.
Structure template:
Problem: X is important because ...
Gap: Existing works fail because ...
Insight: We observe that ...
Approach: We propose ...
Contributions: Idea/Design/Results
Reviewer starts considering acceptance here.
Background + Motivation: Purposeful, selective (not wiki/textbook)
Motivation three-layer structure:
Observation(measurable) → Limitation(inevitable) → Opportunity(actionable)
"Premature solutions kill insights." State "why" first, then "how."
The solution should feel inevitable after reading Motivation.
Structure template:
We observe that existing ...
This is problematic because ...
Existing approaches fail to address ...
This opens an opportunity to ...
Supporting Sections
Design: Explain trade-offs, not describe flow; modular (module-by-module); abstract method over implementation. Maintain separation between Problem Statement (Moti) and Solution Space (Design).
Implementation: Credibility boost, not workload display; keep brief; focus on core parts and key tricks.
Evaluation: Not "running experiments" but verifying propositions:
- Claim 1: faster → latency/throughput
- Claim 2: more efficient → cost/utilization
- Claim 3: more general → workload diversity
- Claim 4: design effective → ablation
Analysis formula:
Result → Cause → Attribution → Implication
Template:
We observe that ...
This is because ...
This validates that ...
This implies ...
Reviewer decides acceptance here.
Related Work: Categorize and contrast; highlight necessity of your design. Three comparison types:
- Fundamentally Different: brief mention
- Orthogonal/Complementary: can combine
- Optimized: quantitative comparison
Conclusion: Fill to full page; can include future work.
Expression Standards
Paragraph Flow
Paragraphs like prayer beads—round, smooth bodies connected by a thread.
Each paragraph: connect up → body → connect down. Uniform length; no isolated short paragraphs.
Cognitive Load Minimization
Writing goal: use simplest logic to help Reviewer understand complex work.
If you can't explain it simply, you don't understand it well enough.
- Every reading obstacle (vague pronouns, long clauses, messy figures) consumes Reviewer goodwill
- If idea/design can't be stated simply, likely not well understood or not well formed
- Complexity in system should NOT translate to complexity in writing
- Say NO to your "code-writing self": delete anything not contributing to core argument
- A good paper is NOT a collection of facts; it is a chain of logically linked arguments
10-Second Rule
If reviewer can't roughly understand a figure in 10 seconds, it's likely a negative asset.
Figures must be self-explanatory; reviewer only viewing figures + captions should understand 80%.
Word Choice
| Avoid | Reason |
|---|
most, best | Extreme unless theoretically proven |
clever, genius | Subjective exaggeration |
precise, optimal | Overclaim risk |
| Inconsistent naming | One term = one concept throughout |
Positive framing: fails to → is limited by → is designed for
Four-Dimensional Evaluation Framework
① Problem Criticality — "Is the problem key?"
| Check | Question |
|---|
| Existence | Is the problem real? |
| Impact | Is it a critical bottleneck? |
| Gap | Is it unsolved? |
→ Reviewer concludes: "This is a real and critical problem."
② Insight Validity — "Is the insight valid?"
| Check | Question |
|---|
| Essence | Does it reveal deep causes behind surface phenomena? |
| Applicability | Is the viewpoint general, not specific? |
| Novelty | Does it challenge common sense or provide new perspective? |
→ Reviewer concludes: "This is a compelling and non-trivial insight."
③ Design Inevitability — "Is the design inevitable?"
| Check | Question |
|---|
| Continuity | Is it a natural derivation from insight? |
| Trade-off | Is there clear trade-off reasoning? |
→ Reviewer concludes: "This is a logical and well-justified design."
④ Evidence Sufficiency — "Is evidence sufficient?"
| Check | Question |
|---|
| Completeness | Is it comprehensive? (SOTA/ablation/sensitivity) |
| Explanation | Is analysis in-depth? (why, not just what) |
→ Reviewer concludes: "The claims are convincingly supported."
Core Principles
Argument-driven 论证驱动
Reviewer-centric 面向评审
Logic-closed 逻辑闭环
Insight-led 洞察主导
Common Rejection Reasons
Papers are rejected not because they are bad, but because they create doubts.
| Reason | Question | Fix |
|---|
| Unclear problem | "Why should I care?" | Clarify where/what/why |
| Weak motivation | "Is this problem real?" | Data-driven support |
| No clear insight | "Just engineering" | Abstract core idea |
| Unjustified design | "Why this design?" | Trade-offs under constraints |
| Weak evaluation | "I don't believe results" | SOTA + ablation + explanation |
| Poor presentation | "Hard to read" | Cognitive load minimization |
| Overclaim | "Exaggerating" | Be modest and rigorous |
Submission Checklist
If any box is unchecked, a reviewer likely will too.
Resources
法无定法，师其意而不泥其形｜No fixed rules—grasp the principle, not the form.
Writing is not a talent — it is a trained discipline.
