sosp-writing-style
GitHub指导SOSP论文写作，强调以设计原则而非功能列表开篇，构建问题到洞察的引言逻辑。提供12页两栏格式的篇幅分配建议，要求图表具备论证力，并确保声明严格受限于实验数据支持。
触发场景
安装
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill sosp-writing-style -g -y
SKILL.md
Frontmatter
{
"name": "sosp-writing-style",
"description": "Use when drafting or revising SOSP prose — leading with a design principle rather than a feature list, building the intro's problem-to-insight arc, budgeting 12 dense two-column pages across design and evaluation, writing figures that argue, and keeping claims inside what the measurements support."
}
SOSP Writing Style
Use this at drafting and revision time for a SOSP paper. The format constants (2026 cycle, checked 2026-07-08 against the official author guidelines): at most 12 pages of technical content in submission, 10-point font on 12-point leading, a 7 x 9 inch text block in two columns, references uncounted. Those pages are read first by a skeptical fast pass and later defended live at a PC meeting — the prose must arm an advocate, not just inform a reader.
Lead with the principle, not the artifact
The recurring identity question at this venue is "what is the principle?" A SOSP paper is remembered as an idea with a system attached, not a system with features attached. Concretely:
- Somewhere on page one, a sentence must exist that survives extraction: the insight, stated so a PC member can repeat it at the meeting without the paper open ("crash consistency can be delegated to the allocation layer because ...").
- Feature-list intros ("we present X, which supports A, B, and C") read as workshop papers. Replace the list with the tension the design resolves.
- If the honest one-sentence version is "we engineered X to be faster", the routing
question comes before the writing question (see
sosp-topic-selection).
The first-page arc
| Move | Job | Symptom when missing |
|---|---|---|
| Problem, concrete | Reader can picture the failing scenario | Intro opens with a market-size claim |
| Why existing systems fail | The gap is structural, not un-tried | Related work reads as the real intro |
| Insight | The extractable principle | Reviewers summarize the paper differently from you |
| Embodiment | The system, named, one paragraph | Insight floats with no evidence anchor |
| Headline evidence | 2-3 measured numbers with baselines named | "Extensive experiments show" |
| Honest scope | One clause on what the design gives up | The meeting discovers the limitation for you |
Budgeting twelve dense pages
A workable allocation for a design-plus-evaluation SOSP paper; deviate deliberately:
Intro + motivation 1.5 pp the arc above, plus the motivating measurement
Design 4.0 pp principle -> mechanisms; invariants stated inline
Implementation 1.0 pp what was built, LoC, what is borrowed vs new
Evaluation 4.0 pp claims mapped to experiments (see sosp-experiments)
Related work 0.75 pp positioning, not survey (see sosp-related-work)
Discussion + conclusion 0.75 pp limitations owned, generality argued
(references uncounted — never starve them)
When over length, cut prose before leading or figures: geometry tampering is a
desk-reject at this venue, and shrunken figures cost more reviewer goodwill than a
trimmed subsection. Move depth (full proofs, extra sweeps) to the supplement, never
review-critical evidence (see sosp-supplementary).
Figures that argue
Systems reviewers read figures before prose on the fast pass. Each figure earns its column-inches by making a claim on its own:
- Caption states the takeaway ("Recovery replays only the log tail; time is flat in log size"), not the axes ("Recovery time vs. log size").
- Baselines appear in the figure, labeled with the same names the text uses.
- Distributions for tail claims: a p99 bar with no spread is an invitation to a soundness attack.
Register and claim discipline
- Present tense, active voice, the system as subject: "Foo batches acknowledgments" beats "acknowledgments are batched by our approach."
- Every superlative is a target at the PC meeting. "Fast" is free; "faster than Bar on write-heavy YCSB by 2.1x" is defensible; "optimal" had better come with a proof.
- Anonymized self-reference in third person throughout — the double-blind rules reach into style ("the Foo paper showed", not "we previously showed").
- Define each invariant once, name it (I1, I2 ...), and cite the name thereafter; the response phase will need those anchors when a review misstates one.
# Weasel-word and overclaim sweep before the freeze
pdftotext main.pdf - | grep -inE 'significant(ly)? (improve|outperform)|extensive experiments|novel(ty)? |clearly|obviously|optimal' | head -20
Revision protocol
- Extract the paper skeleton: first sentence of every paragraph, read in order. If the skeleton does not carry the argument, restructure before polishing.
- Give the draft to a systems reader for a strict 15-minute pass — abstract, intro, figures, conclusion — and ask them to state the principle. Wrong answer means a page-one problem, whatever they say about section 4.
- Only then line-edit for density: merge stub paragraphs, collapse enumerations that repeat their lead-in, delete signposting that restates section titles.
Output format
[Principle] extractable one-liner found on page one? <quote it>
[Arc audit] problem / gap / insight / embodiment / evidence / scope — which missing
[Budget] section page allocation vs the 12-page plan
[Figures] each figure's caption-claim; baselines in-figure?
[Claim sweep] overclaims found -> defensible rewrite
[Skeleton verdict] argument carries on first sentences alone? <yes/no>
版本历史
- 9f86f09 当前 2026-07-19 17:39
