paper-writing

name	paper-writing
description	Generic-purpose writing guide for ML/NLP/DL papers targeting ACL/ICLR/ICML/AAAI/NeurIPS/EMNLP. Distills Widom, Shwartz, SNL-UCSB, and Pengsida into a four-stage pipeline. Use whenever drafting or revising any section. Always run a cohort-research pass against the target venue before drafting.

Paper Writing — A Workflow + Section Guide for ML/NLP Papers

When to use

Trigger on any of:

• "write/draft the X section" (X = abstract / introduction / related work / method / experiments / conclusion)
• "polish/revise/tighten/compress the X section"
• "我想写论文的 X 节" / "帮我改 X 这一节"
• "review my draft" / "审一遍我的写作"
• "我要投 ACL/ICLR/ICML/AAAI/NeurIPS/EMNLP" — any time a target venue is named

Do NOT open LaTeX and start drafting before running the cohort-research pass (Section 2). Generic writing checklists yield generic prose. Venue conventions live in the published cohort, not in this skill.

---

Section 1 — Pipeline at a glance

This skill enforces a four-stage pipeline. Compress freely, but do not skip:

Stage	Output	When	Key references
1. Cohort Research	1-page distillation of target-venue writing conventions	Once per paper, before architecture	Section 2 below; references/cohort-example/distillation.md
2. Architecture	Section outline + claim list + figure/table plan	After cohort research, before drafting	references/workflow.md
3. Drafting	Section drafts in enforced order: Draft 0 Intro → Method → Experiments → Related Work → Final Intro → Abstract → Conclusion	Once architecture is locked	references/sections/<section>.md
4. Compression	30–50% shorter draft	After all sections drafted, before submission	references/compression.md

Cross-stage concerns (style, claim-evidence alignment, conflict resolution) are addressed inline at each stage.

---

Section 2 — Cohort Research (the first stage)

This is the single highest-leverage step.

Why cohort first. What counts as a "Method" vs an "Experiment", what level of abstraction is expected, how modules are named, whether Related Work goes before or after Experiments — these are unwritten conventions of the target subfield. They are NOT in any generic checklist; they live only in the published cohort.

Workflow

1. Identify cohort: 3-5 papers in the target venue / subfield. Must include at least one paper accepted at the EXACT target venue (or its sister venue from the same series — ACL ↔ EMNLP ↔ NAACL; ICLR ↔ ICML ↔ NeurIPS).

2. Add 1 cross-field exemplar: a paper outside the immediate subfield whose section is admired by the user. Prevents cohort tunnel-vision.

3. Pull arxiv LaTeX e-print (NOT PDF):

   curl -sSL https://arxiv.org/e-print/<id> -o paper.tar.gz
   mkdir <name> && tar -xzf paper.tar.gz -C <name>
   

   PDF table-reading is unreliable; always go to source. Cited numbers, table contents, and notation conventions should all be verified against the .tex.

4. Locate the target section in each paper — single-file papers need grep -nE '^\\?section'; multi-file papers usually have sections/<name>.tex.

5. Read across the cohort with a fixed lens. Default lens (extend per section):

   Lens	What to capture
   Opening sentence	Framing device (analogy / unifying concept)? Or dive straight to a definition / problem?
   Subsection naming	Engineering label vs role-name vs design-signal name?
   First sentence of each subsection	Problem statement / role assertion / direct definition?
   Number placement	Where do raw values appear? Where are symbols used? Where is each kind of detail isolated?
   Equation density	Sparse (system style) vs heavy (algorithm style)?
   Justification placement	Separate "Advantages" paragraph vs embedded in design prose?
   Implementation Details	In Method body? End of Method? Experiments? Appendix?
   Voice	Present-tense declarative ("we design") vs procedural ("we configured")?

6. Distill the writing 思路 — qualitative, ≤1 page, plain text. Capture what this section is for in this subfield, not "language patterns".

7. Then draft.

Where to store cohort papers

One cohort folder per paper, not one global cache:

• Default location: <paper-repo-root>/cohort/. For example, if your paper repo lives at ~/my-paper/, your cohort folder should be ~/my-paper/cohort/.
• Layout: one extracted arxiv tarball per subdirectory, e.g., cohort/<paper-name>/. Save the 1-page distillation at the root as cohort/distillation.md.

Why one-per-paper, not one global:

• Prevents cross-domain pollution — observations from one subfield's cohort should not bias a follow-up paper in a different subfield.
• Survives within-domain reuse — the same paper, revisited weeks later, lands in the same folder; no re-downloading.
• Travels with the repo — when the paper is git-tracked or shared, its cohort travels along.

If the user prefers a different location (e.g., outside the repo to keep it untracked), ask once at the start of cohort research and use it consistently for that paper.

What "good distillation" looks like

A worked example is provided at references/cohort-example/distillation.md, with the corresponding raw .tex files in references/cohort-example/method_sections/. This example is from one specific subfield; treat it as a format reference only — do not assume any of its observations transfer to your subfield. Open it, see the structure (per-paper notes + cross-cohort patterns, lensed via Section 2 step 5), then build a fresh distillation from scratch for your own target.

Cohort wins over generic advice

When this skill's generic rules conflict with what the target-venue cohort actually does, the cohort wins. The same principle applies to baseline choice — anchor numbers to what the venue cohort has published, not vendor blog posts or this-month's leaderboard. Writing decisions (Related Work placement, equation density, named-module conventions) follow the same rule.

---

Section 3 — Per-section dispatch

After cohort research, load the relevant section file. Each file contains three pieces (matching the structure validated on Method in the legacy skill):

• Rhetorical moves — what the section is doing, paragraph by paragraph
• Pre-writing template — to fill in plain text before opening LaTeX
• Self-check checklist — to run after drafting, before declaring done

Section	File	Primary template
Title + Abstract	references/sections/title-abstract.md	Widom abstract rules + Shwartz NLP-paper conventions
Introduction	references/sections/introduction.md	Widom five-point as main spine; SNL-UCSB 6-move as cohort variant
Related Work	references/sections/related-work.md	3-move category-cluster sequence; placement decision rule
Method	references/sections/method.md	Framing device + named-module triad + Symbol-Value-Argument boundary
Experiments	references/sections/experiments.md	6-move setup→head-to-head→deep-dive→takeaway→ablation→robustness
Conclusion	references/sections/conclusion.md	Restate-contribution + future-work-marks-territory

Cross-section concerns live in:

• references/workflow.md — writing order, why Intro is written twice, Abstract last, draft-zero principles
• references/style.md — sentence-level + LaTeX/citation conventions (Shwartz + Widom + ML/NLP idioms)
• references/compression.md — SNL-UCSB's 7 compression operations, used in Stage 4

---

Section 4 — Writing order and "Intro twice"

Detailed in references/workflow.md. Key rules:

• Draft 0 Introduction is written before Experiments, as a framing scaffold. It establishes the contribution claims that constrain what Experiments must prove. It is explicitly disposable.
• Final Introduction is rewritten from scratch after Experiments are finalized, constrained by actual evidence. Results Preview (Move 5/6) now contains real numbers.
• Abstract is written last, after Final Introduction. Abstract is a compression of Introduction, not a separate composition.
• Conclusion is written after Abstract — same idea, different audience (the reader who finished vs. the reader deciding to read).
• Related Work in ML/NLP goes after Method/Experiments by default (cohort convention). Override only if your cohort places it earlier.

Source crosswalk:

• SNL-UCSB enforces this ordering explicitly ("Draft 0 Introduction → Evaluation → Design → Background → Related Work → Final Introduction → Abstract").
• Shwartz: "Write introductions last."
• Widom: notes Introduction structure but is silent on ordering. No conflict.

---

Section 5 — Conflict resolution rules

When two sources (Widom / Shwartz / SNL-UCSB / Pengsida / this skill) disagree, apply in order:

1. Cohort wins. If the target-venue cohort (≥3 of 5 recent accepted papers) does X, do X — even if a source says Y.
2. Venue priors. Within ML/NLP top venues, default to:
   • Related Work after main body (not before).
   • First-person plural ("we").
   • Active voice; present tense for method, past tense for experimental procedure.
   • "Large language models" lowercased.
   • \citep for parenthetical, \citet for in-text (NLP) / \cite{} (default; some ICML/NeurIPS templates differ).
   • Number-first contributions in Introduction.
3. Sources rank: Shwartz > Widom > SNL-UCSB > Pengsida for ML/NLP-specific style; Widom > Shwartz > others for structural Introduction guidance (her five-point predates and grounds the others).
4. Skill convention. This skill's defaults below.

Recurring conflicts pre-resolved:

Question	Decision	Why
Where does Related Work go?	After Method/Experiments	ML/NLP cohort default; SNL-UCSB also recommends post-eval; overrides Widom's "depends on length"
When write Introduction?	Twice (Draft 0 then Final)	SNL-UCSB; aligns with Shwartz "write intro last" via the Final pass
Intro structure	Widom five-point	Predates and grounds 6-move; simpler; ML/NLP-canonical
Method opening	Framing device / unifying concept	Cohort-consensus (legacy skill's Section 3 Principle 1); never open with implementation
Numbers in Method	Symbols only, values isolated to Implementation Details / Experiments	Symbol-Value-Argument rule from legacy skill
Capitalization	"large language models" not "Large Language Models"	Shwartz; ACL convention
Hedging	None ("We show", not "We believe")	SNL-UCSB voice rule; ML/NLP cohort default
Revision posture	Forward positive claims, not defensive negation	Reviewer concerns should sharpen what the method is and proves, not make the paper sound apologetic

---

Section 6 — Pre-flight before every drafting session

Before opening LaTeX, confirm:

• Target venue is named (ACL / ICLR / ICML / AAAI / NeurIPS / EMNLP / …).
• Page limit known. Submission deadline known.
• Cohort distillation exists for this paper (run Section 2 if not — save under references/cohort-<subfield>/ for reuse next paper).
• Locked claims list exists: 3-5 contributions, each with one headline number from completed experiments. If experiments aren't done, only Draft 0 Intro is permitted — do NOT draft Final Intro / Abstract / Conclusion yet.
• The section to draft is the next one in the enforced order (Section 4). Drafting Abstract before Final Intro, or Final Intro before Experiments, is a flag.

If any box is unchecked, fix it before continuing.

---

Section 7 — Common pitfalls (ML/NLP-specific)

Drawn from cohort observation and lessons from past submissions. Pause and check before submitting.

• Claim-evidence misalignment — Abstract / Intro promise X, Experiments only show X′ ⊊ X. Every claim in Intro must map to a specific Experiments subsection. Maintain a claim-to-subsection table while drafting; if a claim has no home in Experiments, narrow the claim.
• Vendor leaderboard comparisons — anchor baselines to venue-cohort published numbers, not current-month blog posts. A vendor's reported number on a benchmark is not the same as the venue's cited number for the same model.
• Below-noise-floor A/B claims — single-run differences on noisy benchmarks (memory, long-context, RAG) are unreliable. State seed protocol; use mean±std or paired bootstrap.
• Per-benchmark hyperparameter assumptions — what works on benchmark A may not transfer to B. Either show transfer or scope your claims to the benchmarks you actually tuned on.
• PDF-based reference checking — when verifying a number you cite, pull arxiv e-print LaTeX source and grep the .tex. Do NOT trust PDF table extraction; column alignment and merged cells routinely produce silent misreads.
• Mis-cited baselines — vendor blog numbers ≠ paper numbers; reproductions ≠ published numbers. Decide once which you cite and stay consistent across the main table, Abstract, and Introduction.
• Defensive negation after reviews — revision prose often drifts into "our method is not merely X / does not suffer from Y / is not ad hoc" because it is reacting to reviewer criticism. Rewrite from the positive mechanism forward: "Our method models X via Y, which enables Z." Use negation only after the positive claim is established and only to mark a precise boundary or structural contrast.

---

Section 8 — Source provenance

This skill distills four sources and resolves their conflicts via Section 5.

• Jennifer Widom — "Tips for Writing Technical Papers" (Stanford InfoLab, 2006). https://cs.stanford.edu/people/widom/paper-writing.html. Source of the Introduction five-point structure (Problem / Why important / Why hard / Why not solved / Approach + results) and the Contributions-as-bullet-list convention.
• Vered Shwartz — "Tips for Writing NLP Papers" (Medium, 2023). Source of ML/NLP-specific style: minimalism, "intro answers Why before How", concrete recurring example, capitalization, ACL Anthology citation, \citep vs \citet.
• SNL-UCSB paper-writing-skill (UC Santa Barbara Systems & Networking Lab). https://github.com/SNL-UCSB/paper-writing-skill. Source of: five-stage pipeline, write-Intro-twice, 6-move Intro and 6-move Evaluation sequences, 7 compression operations, voice rules, claim-first headings.
• Pengsida (彭思达) — learning_research, via Master-cai's adaptation (https://github.com/Master-cai/Research-Paper-Writing-Skills). Source of: three-piece module structure (problem → design → advantage), three-level clarity check (logic / paragraph / sentence), sentence skeletons.

The legacy skill paper-writing-legacy retains the memory-systems-specific cohort distillation (LightMem / MemGPT / A-Mem / MemoryOS / Self-RAG / SeCom / LongMemEval / MemAgent / MemoryT1 / BEAM / MemGAS) for reference. The cohort .tex files and distillation are duplicated under references/cohort-example/ here for use as a format example only — do not assume those conventions transfer to other subfields.

Section 9 — Update path

• When working on a paper in a new subfield: run Section 2 cohort research, save the new distillation at references/cohort-<subfield>/distillation.md for reuse.
• When a generic rule in this skill is contradicted by a strong cohort: revise the rule in Section 5 with the cohort evidence, rather than overriding silently.
• When a section file in references/sections/ is found wanting (e.g., draft kept failing checklist), update both the move list and the checklist in the same file — keep the three pieces synchronized.
• Carry-over principles worth re-loading at the start of each drafting session (see also Sections 2, 5, 7):
  • Pull arxiv LaTeX e-print, not PDF, when verifying any cited number.
  • Anchor baselines to the venue cohort, not vendor leaderboards.
  • Once a paper's framing is locked (its claim list and target venue), do not silently widen claims mid-draft — adjust the locked framing first, then the prose.