# Academic Writing

> Write conference papers for systems venues (OSDI, NSDI, SIGCOMM, MOBICOM, SOSP, FAST) and the AAAI AI/ML venue. Organized by venue: pick the paradigm (systems problem-solution-evaluation vs. AAAI/AI-ML method-experiment), then follow venue-specific structure, format, citation style, and review rules. Covers paper structure per venue, writing style, IEEE/ACM numbered vs. AAAI author-year citations, figures/tables, AAAI's mandatory reproducibility checklist and double-blind two-phase review, and FAST short/deployed-systems papers. Use when asked to write, draft, structure, or format a systems or AAAI conference paper.

- Skill: `drunkcoding/academic-writing` (Agent Skill, multi-file: 6 files)
- Install (CLI): `npx skillmds@latest add drunkcoding/academic-writing`
- Raw SKILL.md: https://api.skillmd.com/api/skills/drunkcoding/academic-writing/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Research & Search
- Author: drunkcoding (https://skillmd.com/u/drunkcoding)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/drunkcoding/academic-writing

---


# Conference Paper Writing

This skill covers two paper paradigms across two venue families. Start by identifying the target venue, then follow the matching structure and load the matching reference file.

## Final delivery: humanize the output

**REQUIRED SUB-SKILL (run before returning any draft to the user).** Load the `humanizer` skill (via the skill tool) and run the finished text through it as a final editing pass. It removes AI-writing tells (inflated significance, copula avoidance like "serves as", rule-of-three, overused AI vocabulary such as "delve"/"testament"/"landscape", filler phrases, sycophancy, and em/en-dash overuse) without inventing any fact, name, number, date, or citation. Preserve the format and conventions this skill requires (mandated section headers, citation style, page limits), and if the user supplied a writing sample, pass it so humanizer matches their voice.

## 1. Choose the Venue and Paradigm (start here)

Two paradigms, because a systems paper and an AAAI paper are organized around fundamentally different things:

| | Systems paradigm | AI/ML paradigm (AAAI) |
|---|---|---|
| **Organized around** | A system that was built, deployed, measured | A method/model/theory and the evidence it works |
| **Structure** | Problem → Design → Implementation → Evaluation | Problem → Method → Analysis → Experiments |
| **Detailed reference** | `references/systems_paper_structure.md` | `references/aaai_paper_structure.md` |

### Master venue table

| Venue | Paradigm | Format / template | Tech. pages | Review | Citations | Reproducibility |
|-------|----------|-------------------|-------------|--------|-----------|-----------------|
| **OSDI** | Systems | USENIX | 12 | Single-blind | Numbered (IEEE) | Artifact: strongly encouraged |
| **NSDI** | Systems | USENIX | 12 | Single-blind | Numbered (IEEE) | Artifact: encouraged |
| **SIGCOMM** | Systems | ACM sigconf | 12 | Double-blind | Numbered (ACM) | Artifact: encouraged |
| **MOBICOM** | Systems | ACM sigconf | 15 | Double-blind | Numbered (ACM) | Artifact: encouraged |
| **SOSP** | Systems | ACM sigconf | 15 | Double-blind | Numbered (ACM) | Artifact: required |
| **FAST** | Systems | USENIX | 12 long / 6 short | Double-blind | Numbered (IEEE) | Artifact: encouraged |
| **AAAI** | AI/ML | AAAI Press two-column (`article` + `aaai<NN>.sty`) | 7 (refs + checklist outside) | Double-blind, two-phase | Author-year (natbib) | Checklist: **mandatory in PDF** |

> Numbers drift each year. Confirm page limits, deadlines, and template names against the current edition's official author kit / Call for Papers before submitting.

**Routing:** for a systems venue → read Section 3 and load `references/systems_paper_structure.md`. For AAAI → read Section 4 and load `references/aaai_paper_structure.md`. Writing craft shared by both (principles, citations, figures) is in Section 5.

---

## 2. The Two Paradigms in One Screen

| Aspect | Systems (OSDI/SIGCOMM/FAST) | AAAI / AI-ML |
|--------|-----------------------------|--------------|
| Core contribution | A built system + its engineering | A method/model/theory + scientific evidence |
| Implementation section | Common, explicit (LOC, frameworks) | Not required; folded into Experimental Setup |
| Related Work placement | After Evaluation | Frequently early (convention) |
| Mathematics | Minimal, practical | Central: objectives, proofs, assumptions |
| Evaluation axes | Throughput, latency, scalability, cost | Accuracy/quality, robustness, sample/compute efficiency, provable properties |
| Length | 12-15 pages | 7 technical pages |
| Reproducibility | Artifact evaluation (badges) | Mandatory reproducibility checklist in the PDF |

Everything downstream — structure, section order, what "evaluation" means, citation style — follows from which column you are in.

---

## 3. Systems Conference Papers (OSDI, NSDI, SIGCOMM, MOBICOM, SOSP, FAST)

Standard structure (problem-solution-evaluation). Design + Evaluation together should be ~50% of the paper.

| Section | Pages | Purpose |
|---------|-------|---------|
| Introduction | 1.5-2 | Problem, solution overview, numbered contributions |
| Background/Motivation | 1-2 | Technical context + **empirical evidence** the problem exists |
| Design | 2-3 | Architecture, mechanisms, trade-offs (explain WHY) |
| Implementation | 0.5-1 | LOC, languages, frameworks, integration |
| Evaluation | 2-3 | Macro/micro-benchmarks, ablation, scaling |
| Related Work | 0.5-1 | Differentiation from prior systems (placed after Evaluation) |
| Conclusion | 0.25-0.5 | Summary, no new information |

**Section highlights:**
- **Title**: "SystemName: Descriptive Subtitle." Name the system.
- **Abstract**: unstructured narrative (150-250 words): problem → gap → solution → key numbers → availability. Write last.
- **Introduction**: 6-7 paragraphs — broad context with numbers, technical gap, solution + key insight, 3-4 numbered contributions mapping to sections.
- **Motivation**: must DEMONSTRATE the problem with measurements/profiling, not assert it.
- **Design**: explain WHY, discuss rejected alternatives, address failure modes.
- **Evaluation**: exact hardware/software/workloads/baselines; every introduction claim must be supported here.
- **Related Work**: after Evaluation; organize by category; always differentiate ("Unlike X, we do Y because Z").

**Venue-specific requirements:**

| Venue | Format | Pages | Review | Artifact |
|-------|--------|-------|--------|----------|
| OSDI | USENIX | 12 | Single-blind | Strongly encouraged |
| NSDI | USENIX | 12 | Single-blind | Encouraged |
| SIGCOMM | ACM sigconf | 12 | Double-blind | Encouraged |
| MOBICOM | ACM sigconf | 15 | Double-blind | Encouraged |
| SOSP | ACM sigconf | 15 | Double-blind | Required |
| FAST | USENIX | 12 / 6 | Double-blind | Encouraged |

**Double-blind systems venues (SIGCOMM, MOBICOM, SOSP, FAST):** anonymize the system name, remove repository URLs, cite own work in third person, remove acknowledgments.

**FAST paper categories:**
- *Short papers (6 pages, excl. references):* complete research with full problem statement and evaluation; prefix title "Short Paper: " at submission; same rigor as long papers on a focused contribution — do not sacrifice evaluation for space.
- *Deployed-systems papers:* operational systems with practical deployment lessons; prefix "Deployed System: "; validate with production data; emphasize operational experience, failure modes, lessons at scale.
- *Supplemental material:* optional single PDF, no page limit; reviewers not required to read it.

For full section-by-section guidance, examples, tense guide, and per-venue variations, load `references/systems_paper_structure.md`.

---

## 4. AAAI (AI / ML) Papers

AAAI follows the **method-experiment** paradigm. Conventional structure (strong convention, not an AAAI compliance rule):

Abstract → Introduction → Related Work → Background/Preliminaries → Method → Theory *(if any)* → Experimental Setup → Results → Ablations → Limitations → Conclusion → References → **Reproducibility Checklist**.

**Hard format rules (enforced strictly — violations risk desk rejection):**
- **7 pages of technical content.** References are outside the limit; the reproducibility checklist is appended after references and also does not count. Any in-paper appendix with proofs/experiments IS technical content and must fit in 7 pages — push overflow to the separately uploaded technical appendix.
- **AAAI Press two-column template, required even for anonymous submission.** LaTeX = `article` class + the edition-specific style package (`\usepackage{aaai2026}`), **not** an `aaai.cls`. Do not add `hyperref`/`navigator`.
- **No page numbers/headers/footers.** Body text black; color only in figures (CMYK, grayscale-safe). **Table captions go BELOW the table.**
- Camera-ready: 7 base pages + up to 2 purchasable extra technical pages.

**Structure implications of the 7-page budget:** background and related work are compressed; Method + Experiments (or Method + Theory) form the bulk; there is **no required standalone Implementation section**; AI/ML papers use substantially more mathematics (objectives, proofs, assumptions) than systems papers.

**Reproducibility checklist (mandatory):** completed by every author at submission, placed in the main PDF after the references, shared with reviewers, and factored into the decision. It covers general clarity, theory (assumptions/proofs), datasets (appropriateness/availability), and computation (seeds, hardware/software versions, number of runs, variance, significance tests, hyperparameters/search ranges). Write the paper so every item can honestly be answered "yes."

**Review process:** double-blind (cite own work in third person; preprints/arXiv allowed but not cited/linked from the anonymous paper); mandatory **abstract registration** ~a week before the paper deadline; **two-phase review** where Phase-1 rejects get no feedback; a single author-feedback/rebuttal window for survivors (AAAI-26: 2,500-char cap, no URLs/files/new results). No dual/concurrent archival submission.

> There is **no separate "CARE" checklist** and **no official fixed reference count** for AAAI — do not invent either.

For full section-by-section guidance, the complete checklist breakdown, appendix mechanics, and the systems↔AAAI contrast, load `references/aaai_paper_structure.md`.

---

## 5. Shared Craft (both paradigms)

### 5.1 Writing principles

- **Clarity**: precise technical language; define terms at first use; active voice ("We design", "We propose", "We evaluate").
- **Conciseness**: strict page limits; every sentence earns its space; ~15-20 word sentences.
- **Directness**: state claims with numbers ("2.3x higher throughput", "+4.1 points accuracy"); avoid hype ("novel", "groundbreaking").
- **Precision**: exact configurations; consistent units and metrics; distinguish observation from interpretation; report variance.

Full guidance, revision checklists, and the systems-vs-ML style comparison: `references/writing_principles.md`.

### 5.2 Citations

- **Systems venues** → numbered citations `[1]`: IEEE (USENIX: OSDI/NSDI/FAST) or ACM (SIGCOMM/MOBICOM/SOSP). Cite systems by name ("Borg [5]").
- **AAAI** → author-year `(Smith 2025)` via `natbib` (`\citet`/`\citep`/`\citeauthor`/`\citeyear`); the AAAI `.sty` sets the matching `.bst`.
- Cite recent work heavily; cite every baseline you compare against; keep self-citation modest and third-person in double-blind submissions.

Full formats, BibTeX, and the numbered↔author-year comparison: `references/citation_styles.md`.

### 5.3 Figures and tables

- **Tables** for exact numbers, configurations, and comparisons; **figures** for trends, distributions (CDFs), scaling, and architecture/method diagrams.
- Self-explanatory captions; error bars on measurements; **bold the best result**; consistent colors; grayscale-safe.
- AAAI-specific: table captions **below** tables; figures ≥300 dpi in `.pdf`/`.png`/`.jpg` (no EPS/GIF), no Type 3 fonts.

Full guidance and per-venue requirements: `references/figures_tables.md`.

### 5.4 Terminology

Use precise subfield terminology. Systems: consistency models, fault tolerance, scheduling, memory hierarchy, RDMA/DPDK/eBPF, parallelism (data/model/pipeline/tensor), TTFT, p50/p99. AI/ML: objective/loss, optimization, generalization, ablation, sample/compute efficiency, seeds/variance, assumptions/propositions/proofs. Define abbreviations at first use; use standard ones (GPU, API, SGD) without definition.

---

## 6. Common Pitfalls

**Both paradigms:**
- Passive voice obscuring who did what; vague quantification ("significant" vs. a number); overstated claims not supported by results; inconsistent terminology; abstract that is a prose table of contents.

**Systems-specific rejection reasons:** not solving a real problem (weak motivation, no evidence); not actually solving it (evaluation doesn't support claims); missing/unfair baselines; unjustified design trade-offs; incomplete related work.

**AAAI-specific pitfalls:** exceeding 7 pages or hiding technical content in "reference" pages; formatting violations (wrong template, added page numbers, `hyperref`); single-seed point estimates without variance/significance; a reproducibility checklist the paper cannot honestly support; de-anonymizing via first-person self-citation or arXiv links.

---

## 7. Workflow

**Systems paper drafting order:**
1. Design + architecture figures → 2. Evaluation + result figures → 3. Introduction → 4. Background/Motivation → 5. Implementation → 6. Related Work → 7. Conclusion → 8. Abstract → 9. Title.

**AAAI paper drafting order:**
1. Method + core equations/pseudocode → 2. Experiments (setup, results, ablations) *or* Theory → 3. Introduction → 4. Related Work + Preliminaries → 5. Limitations + Conclusion → 6. Abstract → 7. Title → 8. Reproducibility checklist (fill as experiments finalize; if an item can't be "yes", fix the paper, not the answer).

**Shared revision + final prep (both):**
1. Verify every introduction/abstract claim is supported by results.
2. Check all figures/tables are referenced and numbers match across text/tables/figures.
3. Ensure consistent terminology and correct citation style for the venue.
4. Format with the exact venue template; check page limit and figure readability (grayscale).
5. Anonymize for double-blind venues (systems double-blind venues and all of AAAI).
6. Systems: prepare the artifact. AAAI: complete the reproducibility checklist and any uploaded appendix.

---

## References

Load the reference that matches your venue and task:

- `references/systems_paper_structure.md`: Full systems paper structure (OSDI/NSDI/SIGCOMM/MOBICOM/SOSP/FAST) — section-by-section content, examples, tense guide, artifact evaluation, per-venue variations.
- `references/aaai_paper_structure.md`: Full AAAI (AI/ML) paper structure — format/length hard rules, section-by-section content, the mandatory reproducibility checklist, double-blind two-phase review, appendix mechanics, and the systems↔AAAI contrast.
- `references/citation_styles.md`: IEEE and ACM numbered styles (systems) and AAAI author-year (natbib), BibTeX management, and anonymization.
- `references/figures_tables.md`: Architecture/method diagrams, performance plots, CDFs, evaluation tables, visual design, and per-venue figure requirements.
- `references/writing_principles.md`: Core writing principles (clarity, conciseness, accuracy), systems-vs-ML style, and revision checklists.