Version 3.6.0. Customized from Orchestra Research's AI Research Skills library, with writing-framework ideas adapted from NC Manuscript Craft. Write complete supported drafts or revise mature manuscripts, with dedicated guides for abstracts, introductions, related work, methods, experiments and conclusions, plus substantive discussion and source-aware consistency checks. Retains ML-specific theory, algorithms, evaluation and method-figure capabilities. See NOTICE and references/sources.md for attribution.
Draft and prepare ML / AI / Systems papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM, OSDI, NSDI, ASPLOS, SOSP, checking actual venue/year requirements when relevant. Use for research-repository drafting, abstracts, introductions, discussion, scientific revision, citations, method figures and format conversion. Deliverable readiness depends on the evidence and checks actually completed.
请把 https://github.com/Hoemr/ml-paper-writing 下载下来，把完整技能目录（含
SKILL.md、references/、scripts/、templates/、LICENSE、NOTICE）复制进你读取 skills 的目录下的
ml-paper-writing/ 子目录（例如 ~/.claude/skills/、~/.codex/skills/、
~/.zcode/skills/，以你实际使用的平台为准）。完成后读取其中的 SKILL.md
确认可用。不要向我提问，直接做完并汇报安装到了哪个路径。
Download https://github.com/Hoemr/ml-paper-writing and copy the complete skill,
including SKILL.md, references/, scripts/, templates/, LICENSE and NOTICE, into ml-paper-writing/ under the skills directory your agent
reads (for example ~/.claude/skills/, ~/.codex/skills/ or ~/.zcode/skills/ - whichever
applies to your platform). Then read SKILL.md to confirm. Do not ask questions; finish
and report the install path.
| Addition | Where | Purpose |
|---|---|---|
| Task and scope | SKILL.md |
Draft, revise, assess or prepare artifacts at partial or full scope. Complete drafting is explicit; an outline is intermediate when prose was requested. |
| Shared writing standard | SKILL.md |
Integrates anti-defensive-writing-en in the core: center the strongest supported value, avoid work-report chronology and self-undermining, and apply its narrative priorities across all sections. Takes precedence over older section-level advice. |
| Scoped delivery audit | SKILL.md |
Maintain actual coverage, lint applicable LaTeX, review meaning and report unresolved checks. Full drafts receive a gate table; small edits receive a concise verification note. |
| Abstract and introduction | references/abstract-and-title.md, references/introduction.md |
Choose decisive evidence and connect problem, prior capabilities, precise gap and supported advance. |
| Related work, method, experiments, conclusion | Dedicated references in the route table | Full section-level writing guidance, worked examples and completion checks; read only the requested section. |
| Discussion and evidence dependencies | references/writing-guide.md, references/evidence-workflow.md |
Interpret findings and update affected prose, equations, algorithms, captions and appendix when sources change. |
| Mechanical style lint | scripts/lint_style.py |
10 checks with flag/triage severities: manual \ref prefixes vs \cref, lowercase table headers, named artifacts without a citation, over-claim rhetoric (novel, "extensive experiments", "state of the art", …), semicolon chains / explanatory-dash density, single-use macros, --bib missing cite keys, --log overfull & undefined references. |
| Punctuation & notation gates | references/style-conventions.md |
Semicolons and explanatory dashes used sparingly; no single-use intermediate symbols; one declared notation scheme; \cref with abbreviated names where the venue allows. |
| Progressive disclosure | references/ |
Each topic has a designated owner. Abstract, introduction and evidence guidance load only when relevant; artifact-specific mechanics stay in existing modules. |
| Built-in method figures | references/figure-workflow.md, references/method-figure/, scripts/method-figure/ |
Architecture diagrams, method overviews, pipeline schematics, reference-led restyling, collision repair and figure audits, with standalone TikZ, PDF/PNG and six-stage QA. |
| Venue template snapshots | templates/ |
LaTeX snapshots for the 10 supported venues, with per-venue setup notes in templates/README.md. |
- Local edits: verify new/affected references and their support.
- Full drafts/revisions/full-skill runs: audit all bibliography entries and every citation occurrence, including existing references. Report incomplete checks.
- Reference-only audits: report findings without editing the manuscript or bibliography.
- Title/abstract retrieval optimization: explicit request only. Normal title polish and full-skill requests do not run GEO, query coverage, trend or visibility probes.
Examples: “用 ml-paper-writing 全量修订” includes reference auditing; “只核查引用，不修改论文” runs a read-only audit; “启用标题检索优化 workflow” enables the optional title optimizer.
The optional title optimizer now prioritizes semantically appropriate terms with verified recent citation growth: rolling 365-day comparison, fixed paper cohorts, base counts, absolute/relative growth and single-paper concentration. It offers a concise title and an evidence-backed keyword variant, without promising search gains. See references/title-optimization/citation-momentum.md. Publication trends remain separately labeled auxiliary evidence. No live keyword ranking is bundled.
Full drafting and substantive revision now use an integrated argument-design sequence, expanded research-type routes and evidence-specific claim calibration. The existing evidence map links claims, methods/proofs and empirical evidence; figure roles reuse it. These writing principles run by default within scope.
Best-Paper-Writing-Skill contributes optional function/type-indexed exemplars and descriptive style metrics, not its original hard-error checker or mandatory intake/report workflow. See resource guide. Statistics from 55 award papers are reference distributions, not requirements for ACL, journals, theory papers or Chinese prose. Existing figure QA, full-scope reference audits and explicit-only title optimization retain their behavior.
ml-paper-writing/
├── SKILL.md ← shared writing standard, modes, workflow, gates
├── references/
│ ├── writing-guide.md ← complete drafting, research types and sections
│ ├── narrative-design.md ← argument design from evidence to prose
│ ├── research-routes.md ← L1–L8 plus systems and survey routes
│ ├── claim-calibration.md ← strongest wording by evidence type
│ ├── bestpaper/ ← optional exemplars, corpus statistics and licenses
│ ├── abstract-and-title.md ← selection, argument, compression and titles
│ ├── introduction.md ← capabilities, gap, idea and fulfilled promises
│ ├── related-work.md ← source relationships and exact comparison
│ ├── method-writing.md ← idea, equations, procedure and reproduction
│ ├── experiments-results.md ← evaluation questions and result narration
│ ├── conclusion.md ← supported answer and consequence
│ ├── evidence-workflow.md ← sources, dependencies and synchronization
│ ├── style-conventions.md ← notation, cross-refs, tables, prose, punctuation
│ ├── algorithms.md ← pseudocode recipes and worked examples
│ ├── citation-workflow.md ← unified citation verification and audit routing
│ ├── reference-audit.md ← complete identity and occurrence-level support audit
│ ├── title-optimization.md ← opt-in retrieval optimization workflow
│ ├── title-optimization/ ← supporting guidance, snapshots and license
│ ├── figure-workflow.md ← built-in narrative, TikZ construction and QA
│ ├── method-figure/ ← drawing guides, TikZ templates and reference PNG
│ ├── reviewer-guidelines.md ← supporting wording examples, equivalence check
│ ├── checklists.md ← submission rules; verify exact venue and year
│ ├── systems-conferences.md ← OSDI / NSDI / ASPLOS / SOSP specifics
│ └── sources.md ← guidance provenance and external limits
├── scripts/
│ ├── lint_style.py ← mechanical delivery-audit lint (see below)
│ ├── bestpaper/ ← optional exemplar retrieval and style measurements
│ ├── check-ref/ ← literature lookup helper and private local configuration
│ ├── paper-title-optimizer/ ← optional trend, title-statistics and visibility helpers
│ ├── audit_pseudocode.py ← mechanical audit for a pseudocode block
│ └── method-figure/ ← QA runner, lint, instrumentation, geometry/type checks
├── templates/ ← venue template snapshots + setup notes
├── LICENSE ← MIT (from upstream)
├── NOTICE ← upstream and adapted-guidance notices
└── README.md
1. Record the requested scope and directly affected components
2. For edited LaTeX: python3 scripts/lint_style.py paper.tex --bib refs.bib --log main.log
→ triage every finding (fix it, or state why it stands)
3. Fresh-context reviewer (rules + text + needed evidence) when available and permitted;
otherwise a separately identified self-review. Method figures retain independent visual QA.
4. Full-paper work: gate table. Small edits: concise verification note. Report actual gaps.
Checks follow the actual artifact. Plain-text prose needs no artificial compile; a compiled PDF still needs visual review. Never equate lint success with scientific correctness.
- One central contribution with an explicit problem, mechanism, evidence, and consequence.
- Preserve scientific meaning: assumptions, comparison budgets, denominators, uncertainty, adverse results, limitations. Never invent results or citations.
- Default to partial scope for local edits; complete-draft and whole-manuscript requests use full scope.
- Run the delivery audit before reporting completion; report actual coverage and failures.
- Verify current venue rules from official sources — page limits, disclosure policies, and templates change.
The original base remains MIT; see LICENSE. Adapted Best-Paper guidance, scripts and corpus material retain CC BY 4.0 attribution under NOTICE and the resource license. Quoted paper excerpts retain their authors' rights and are not relicensed. See resource provenance.
- Orchestra Research's AI Research Skills library — upstream
- Built-in method-figure module — standalone TikZ source, PDF/PNG, design rules and mechanical/visual QA; no separate figure skill required
- Built-in reference audit — full coverage during full-paper work; read-only when requested as an audit
- Optional title retrieval optimization — explicit opt-in only, excluded even from generic full-skill runs
- Hoemr/my_paper_hub — personal reference paper hub