AGENT SKILL

academic-paper-generation

暂无简介

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 25
Forks
4

02 资源文件 1

03 SKILL.md 正文

Academic Paper Generation: From Concepts to Production-Ready PDF

Scope: Complete workflow for generating publication-ready academic papers
Audience: Academic writing, arXiv submissions, conference papers
Complexity: High (multi-phase orchestration)
Time Investment: Significant initial learning, fast payoff on repeated use


Overview

This skill provides a systematic, reusable workflow for transforming raw technical concepts, architectural decisions (ADRs), and research findings into a complete, polished academic paper suitable for arXiv, IEEE, ACM, or other venues.

The workflow enforces quality gates at each phase to catch common academic writing failures early:

  • Weak narrative arc (no clear story)
  • Thin content (placeholders instead of depth)
  • Broken references (placeholder citations without research)
  • Poor figure quality (artifacts, misalignment, missing context)
  • Structural inconsistency (section depth varies)

The Five-Phase Workflow

Phase 1: Research & Reference Collection (Narrative Planning)

Goal: Establish the paper's core claim and gather supporting references.

Steps:

  1. Read source material (ADRs, concepts, existing docs) and extract the central insight—the one-sentence claim the paper argues.

    • Good claim: "Compliance mechanisms must be structural invariants, not configurable features."
    • Bad claim: "We implemented twelve compliance mechanisms."
  2. Identify the problem space via three historical examples (real incidents, not invented):

    • What went wrong?
    • Why did it happen? (root cause)
    • What does it reveal about current practice?
  3. Dispatch reference research agents (parallel work):

    • Agent A: "Find 15 academic papers on [topic + regulatory domain]"
    • Agent B: "Find 5 industry incident postmortems (Equifax, Facebook, etc.)"
    • Agent C: "Find 10 standards/regulations (GDPR, EU AI Act, HIPAA, PCI-DSS)"

    Each agent returns curated list with URLs, citations, and 1-line summary of relevance.

  4. Build a reference database (CSV or JSON):

    category | author | year | title | url | relevance | sections_to_cite
    ----------|--------|------|-------|-----|-----------|------------------
    regulatory | EU | 2024 | EU AI Act | ... | Article 50 | intro, m1
    incident | Rouse | 2017 | Equifax | ... | compliance failure | intro
    academic | Emerson | 2008 | Model Checking | ... | formal methods | methodology
    
  5. Map references to paper sections (pre-commitment):

    • Introduction (3-5 foundational + 2-3 incident refs)
    • Methodology (3-4 formal methods refs)
    • Each mechanism (1-2 regulatory + 1-2 domain-specific refs)
    • Related Work (5-7 competitive/complementary papers)

Deliverable: A reference database and a one-paragraph problem statement that explains the paper's narrative arc.

Quality Gate: Every section has ≥3 references allocated. No section is empty.


Phase 2: Narrative Arc Design (Story Structure)

Goal: Create a clear, compelling story that carries the reader from problem through solution to validation.

The IMRAD+C Structure (adapted for this domain):

  1. Introduction (40% of body text)

    • Hook (problem statement, 3 incidents showing pattern)
    • Gap (current approaches fail because...)
    • Thesis (our solution: structural invariants)
    • Roadmap (in this paper we...)
  2. Methodology (15% of body text)

    • Formal model (state machines, CTL*)
    • Threat model (what could break compliance)
    • Design principles (fail-closed, audit-first, etc.)
  3. Results / Mechanisms (30% of body text)

    • The twelve mechanisms, each:
      • Regulatory requirement (cite the article)
      • Implementation (how we enforce it)
      • Why it's structural (can't be toggled away)
      • Verification (how we prove it works)
  4. Application / Deployment (10% of body text)

    • CorvinOS as proof-of-concept
    • Production stats (10M+ events, 50+ tenants, 100% audit clean)
    • Performance overhead (measured, not claimed)
  5. Conclusion (5% of body text)

    • Restatement of thesis
    • Implications (compliance as architecture, not configuration)
    • Future work (formal verification, distributed systems)

Narrative Arc Check:

Problem:  "Compliance keeps getting disabled under operational pressure" 
          ✓ Concrete examples (Equifax, Facebook, SolarWinds)
          ✓ Root cause explained (treated as configurable feature)

Solution: "Make compliance a structural invariant instead"
          ✓ Formal definition
          ✓ Twelve orthogonal mechanisms
          ✓ Can't be disabled without system restart

Proof:    "CorvinOS shows this works in production"
          ✓ 12 months, 50+ tenants, zero incidents
          ✓ Measurable audit overhead
          ✓ Real deployment data

Deliverable: A one-page outline with section titles, estimated word count per section, and narrative flow check.

Quality Gate: Reader can understand the paper's claim by reading only the section headings. Every section advances the narrative.


Phase 3: Content Generation (Writing with Depth)

Goal: Write sections with sufficient depth (3-5 well-developed paragraphs, not 1-2 sentences).

Heuristics for Content Density:

Section Length Paragraphs References Formulas
Abstract 150-200 words 1 0 0
Introduction 800-1200 words 4-6 5-8 1-2
Methodology 600-1000 words 3-5 3-5 3-5
Each Mechanism 200-300 words 2-3 1-2 0-1
Related Work 400-600 words 3-4 7-10 0
Conclusion 300-400 words 2-3 0-2 0

Writing Patterns:

For each mechanism (M1-M12), follow this template:

**Regulatory Basis:** "GDPR Article X requires..."
→ Cite the regulation, explain what it's trying to protect

**Current Practice Fails:** "Most systems implement this as..."
→ Show why typical implementations are configurable/optional

**Structural Solution:** "In CorvinOS, we enforce this by..."
→ Explain the architectural decision
→ Show why it cannot be disabled

**Verification:** "Auditors can verify by..."
→ Concrete, measurable verification method
→ Complexity (O(n), O(1), etc.)

**Example Impact:** "When [scenario], the system [response]"
→ Concrete scenario showing the invariant holding

References Within Text:

  • Cite regulations by article number: "GDPR Article 30 requires..."
  • Cite academic work early in sections: "Prior work on compliance-as-code [Ref] assumes..."
  • Cite incidents in introduction: "Equifax (2017) disabled audit logging [Ref]..."
  • Do NOT use placeholder citations like \cite{ComplianceAsCode2019}; research the paper first.

Deliverable: Complete draft of all sections with proper citations embedded.

Quality Gate:

  • No \cite{} placeholders—every citation is researched and valid
  • Minimum 2-3 paragraphs per mechanism
  • Every regulatory requirement is cited to the actual regulation
  • Abstract and conclusion are polished and final

Phase 4: Figures & Diagrams (Quality Visual Communication)

Goal: Create publication-quality diagrams that support the narrative.

Diagram Types & Creation Strategy:

Diagram Type Tools Format Checklist
Threat Model Flow Mermaid / Graphviz SVG → PDF (not PNG) Labels clear, no overlaps
Hash Chain Timeline Sequence TikZ / draw.io PDF Time axis labeled, colors distinct
Compliance Matrix Table TikZ / Excel → TikZ PDF Green=compliant, labels aligned
Multi-tenant Isolation Architecture Mermaid / draw.io PDF Boundaries clear, isolation obvious
Mechanism Overview Infographic Custom SVG PDF Hierarchical layout, legend

Critical Rules for Figure Quality:

  1. SVG → PDF Pipeline (not PNG):

    • SVG source files stored in figures/ directory
    • Use inkscape --export-type=pdf to convert (no rsvg-convert issues)
    • Verify PDF in Acrobat; check for text rendering issues
    • PNG export is only fallback if PDF fails
  2. Caption & Labeling:

    • Caption format: "Figure N: [One-line claim]. [Supporting detail]. [How it relates to narrative]."
    • Example: "Figure 2: Boot-Time Tripwire Sequence. Audit chain is verified before plugins load, ensuring no instance runs with corrupted records. This structural enforcement prevents compliance drift over time."
    • All labels in figures must be English, sans-serif font (Arial, Helvetica)
    • No diagrams use red/green alone (colorblind-safe palette)
  3. Placement & Integration:

    • Figures referenced in text before appearing: "As shown in Figure 2..."
    • No floating figures; every figure has a cite/reference in text
    • Maximum one figure per subsection; typically 1-2 per major section
    • Total figures: 3-5 for a short paper, 8-12 for a long paper
  4. Anti-patterns to Avoid:

    • No screenshots of code (use verbatim blocks instead)
    • No auto-generated diagrams without manual cleanup
    • No figures with blurry text or pixel artifacts
    • No diagrams that replicate content in the text (figures should add, not duplicate)

Deliverable: 4-6 publication-quality diagrams (PDF format) with captions and references integrated into text.

Quality Gate:

  • Every diagram is in PDF, not PNG
  • Every diagram has a caption explaining narrative relevance
  • Every diagram is referenced by text (not orphaned)
  • No text rendering issues; all labels are crisp

Phase 5: LaTeX Assembly & Compilation (Final Production)

Goal: Produce a final, polished PDF suitable for submission.

Template & Structure:

\documentclass[11pt,a4paper]{article}
\usepackage{arxiv}
\usepackage[utf8]{inputenc}
\usepackage{hyperref}
\usepackage{amsmath}
\usepackage{graphicx}

\title{Title}
\author{Author1 \and Author2}
\date{Date}

\begin{document}
\maketitle
\begin{abstract}...\end{abstract}
\section{Introduction}...\end{section}
\section{Methodology}...\end{section}
\section{Results}...\end{section}
\section{Conclusion}...\end{section}
\begin{thebibliography}{99}...\end{thebibliography}
\end{document}

BibTeX Best Practices:

@article{AuthorYear,
  author = "First Last and Second Last",
  title = "Full Title with Capitals",
  journal = "Journal Name",
  year = "2024",
  volume = "10",
  pages = "123--456"
}

@misc{Regulation2024,
  title = "Regulation Title",
  author = "Organization",
  year = "2024",
  note = "Official Journal Reference"
}

@techreport{CISAReport,
  author = "CISA",
  title = "Full Title",
  year = "2020",
  note = "URL or Report Number"
}

Compilation Workflow:

# Step 1: Check for undefined references
tectonic paper.tex 2>&1 | grep -E "undefined|Missing"

# Step 2: Verify figure paths
for fig in figures/*.pdf; do
  pdfinfo "$fig" > /dev/null || echo "✗ $fig is corrupted"
done

# Step 3: Full compilation
tectonic paper.tex --keep-logs

# Step 4: Verify output
pdfinfo paper.pdf | head -15

Final Checks Before Submission:

  • [ ] Abstract is 150-200 words, standalone (no references to sections)
  • [ ] All figures are in PDF, not PNG
  • [ ] No \cite{} placeholders (every citation resolved)
  • [ ] Page count is reasonable (4-8 pages for short paper, 12-20 for full paper)
  • [ ] Margins are 1 inch (0.75-1 inch is acceptable)
  • [ ] All figures have captions and are referenced in text
  • [ ] Keywords are listed (typically 5-7)
  • [ ] No typos in author names or affiliations
  • [ ] References are sorted consistently (alphabetical or numeric)

Deliverable: A final paper.pdf, production-ready and suitable for arXiv/journal submission.

Quality Gate:

  • PDF compiles without errors or warnings
  • All figures render correctly (pdfinfo validates)
  • Page count is reasonable
  • All references are resolved
  • Visual inspection passes (no typos, formatting is clean)

Anti-Patterns to Avoid

Anti-Pattern Why It Fails Fix
Placeholder citations (\cite{ComplianceAsCode2019} without research) Reader can't verify; looks unfinished Research first, cite real papers with URLs
One-sentence mechanism descriptions Lacks depth; looks like a checklist, not analysis 2-3 paragraphs per mechanism; regulatory basis + implementation + why structural
PNG figures with artifacts Blurry, unprofessional; looks auto-generated Use SVG → PDF pipeline; verify in Acrobat
No narrative arc (just lists facts) Reader doesn't understand why they should care IMRAD structure: problem → solution → proof → implications
Figures without captions (self-explanatory claim) Impossible to reference; violates journal guidelines Every figure gets a caption explaining narrative relevance
Section depth varies wildly (1 para, then 5 paras) Looks inconsistent and unprofessional Target density: 3-5 paras per major section
Related work crammed into conclusion Makes conclusion weak; related work hard to find Dedicated Related Work section (one major section)
No production data (only theory) Claims unsupported; looks academic but not grounded Include real deployment stats (page count, latency, audit events)

Implementation Checklist

Use this checklist to track progress through all five phases:

  • [ ] Phase 1: Research

    • [ ] Central claim identified and phrased
    • [ ] 3+ historical examples (incidents) researched
    • [ ] Reference database built (40+ sources)
    • [ ] References mapped to sections
  • [ ] Phase 2: Narrative

    • [ ] One-page outline written
    • [ ] IMRAD structure mapped
    • [ ] Narrative arc checked (problem → solution → proof)
    • [ ] Section word counts estimated
  • [ ] Phase 3: Content

    • [ ] All sections written (no placeholders)
    • [ ] 3-5 paragraphs per major section
    • [ ] All citations researched (no \cite{} placeholders)
    • [ ] Examples and scenarios concrete (not hypothetical)
  • [ ] Phase 4: Figures

    • [ ] 4-6 diagrams created (SVG source files saved)
    • [ ] All diagrams converted to PDF
    • [ ] Captions written and embedded
    • [ ] Figures referenced in text
  • [ ] Phase 5: Assembly

    • [ ] LaTeX template created
    • [ ] BibTeX references formatted correctly
    • [ ] Paper compiles without errors
    • [ ] Final checks passed

Automation Opportunities

This workflow can be partially automated:

  1. Phase 1 (Research): Dispatch three parallel research agents; aggregate results into reference DB
  2. Phase 2 (Narrative): Use template-driven outline generation; validate structure via checklist
  3. Phase 3 (Content): Use generation agents with narrative-aware prompts; verify density heuristics
  4. Phase 4 (Figures): Generate diagram specs from mechanisms; batch-convert SVG → PDF via Inkscape
  5. Phase 5 (Assembly): Template-driven LaTeX generation; automated compilation and validation

Next Skill: "Academic Paper Generator" (orchestrates all five phases via multi-agent workflow)


Resources & References

  • ArXiv Best Practices: https://arxiv.org/help/submit
  • Henning, A. (2020). "How to Write a Good Research Paper." IEEE Computer Society
  • Academic Writing Guides: Purdue OWL (online writing lab)
  • Diagram Tools: Mermaid, Graphviz, TikZ, draw.io
  • Compilation: Tectonic (robust TeX engine)

Key Takeaway

Compliance-as-Architecture, not Configuration applies to papers too: structure the narrative arc, depth, figures, and citations as load-bearing elements of the paper, not as optional flourishes added at the end. A paper with a clear story, sufficient depth, and strong visuals is far more likely to be read, cited, and accepted than one where these elements are afterthoughts.

04 同仓库更多 Skills