Best Skills from Prof. Bingsheng He's daily research experience — packaged as a Claude Code plugin.
The centerpiece skill. It coaches you to write and improve a CS/ML paper (not just critique a finished one), with five modes:
| Mode | What it does |
|---|---|
outline |
Turn your one-sentence contribution into a paper skeleton (story spine, refutable contribution list, section + page budget, the 5 planned figures) |
section |
Write/improve one section against explicit criteria, using verbatim worked examples from Prof. He's real papers |
polish |
Line-level clarity pass (active voice, one-term-per-concept, figures that stand alone) |
checklist |
A single consolidated pre-submission gate (contribution, structure, figures, evidence, reproducibility, mechanics) |
reviewer |
Adversarial "how would Reviewer 2 attack this?" — a 3-minute-skim verdict with the top rejection risks |
Why it's better than existing open-source paper-writing skills: it is the only one that is example-grounded (real before→after from his papers), criteria-driven (an extended XtraGPT taxonomy — his own group's — with the Methods / Related Work / Limitations gaps filled in), venue-aware (DB/systems vs. ML conventions), reviewer-eyed, and lightweight (one skill + reference files, no external APIs) — all at once.
/openbs-paper outline "a scheduler that cuts agentic-flow latency under uneven load"
/openbs-paper section abstract path/to/draft.tex
/openbs-paper reviewer path/to/draft.pdf
The core research workflow chains three skills: /openbs-idea (find an idea) → /openbs-idea-vet (validate & consolidate it) → /openbs-paper (write it up).
| Skill | Description |
|---|---|
/openbs-paper |
Flagship — write & improve a paper: outline, section-by-section drafting against criteria, polish, pre-submission checklist, and an adversarial reviewer pass |
/openbs-idea-vet |
Rapidly validate & consolidate a rough early-stage idea: sharpen it into an "Idea Card", then judge novelty (by actually searching related work and measuring the delta vs. the nearest prior work) and technical contribution (real research vs. engineering), de-risk with the cheapest tests + a timeboxed checkpoint plan, and give a Go / Refine / Strategic-Withdrawal verdict |
/openbs-idea |
Research ideation, PhD Year 1 survival guide, reading strategies, and building good habits |
/openbs-review |
Review paper drafts using Prof. He's quality checklist, 5-figure rule, and 28-point pre-submission checklist |
/openbs-submission |
Paper submission guidelines: double-blind compliance, COI declaration, timeline, and integrity rules |
/openbs-rebuttal |
Write conference paper rebuttals with structured response strategies |
/openbs-camera-ready |
Camera-ready preparation: plagiarism checks, formatting, references, and 18 automated checks |
/openbs-thesis |
Guide QE reports, Thesis Proposals, and Final Theses |
/openbs-presentation |
Prepare conference talks, seminars, thesis defenses, and posters |
/openbs-conference |
Conference attendance prep with the 5-5-5 networking framework |
/openbs-collaboration |
Research collaboration, industry partnerships, and intern management |
/openbs-job-search |
Academic job search: applications, statements, job talks, interviews, and negotiation |
/openbs-broader-impact |
Write Broader Impact sections for NeurIPS and other venues |
/openbs-ai-era |
Use AI/LLMs in research the "augment, don't automate" way — per-phase guidance, recommended prompts, and integrity red-lines |
/openbs-paper and the other craft skills (/openbs-review, /openbs-idea, /openbs-thesis, /openbs-presentation, /openbs-rebuttal) are grounded in verbatim worked examples from Prof. He's own papers — e.g. the 5-move Mars abstract, the "First/Second/Third" one-to-one contributions form, the ablation-pair figure trick, and honest baseline ladders — rather than abstract rules alone. The full library lives in plugins/openbs/references/:
paper-craft-exemplars.md— writing craft (abstracts, intros, the 5-figure rule, design, experiments, trajectory) drawn from his SIGMOD/PACT/TODS/SC papers, with source URLs.section-criteria.md— the per-section revision criteria: the XtraGPT 20-criteria taxonomy (from his own group) extended with the Methods / Related Work / Limitations criteria it omits, plus venue conventions.external-writing-canon.md— a deduplicated, source-tagged distillation of the community canon: Simon Peyton Jones, Whitesides, Widom, Ernst, Levin & Redell, Kording & Mensh, the Heilmeier Catechism, the NeurIPS author checklist, Swales/CARS, and ABT.stories-and-philosophy.md— the war stories, metaphors, mottos, mentorship voice, and AI-era stance behind the advice.
- Add the marketplace:
/plugin marketplace add medusa-compute/OpenBS
- Install the plugin:
/plugin install openbs@openbs
Clone the repo and load the plugin directly:
git clone https://github.com/medusa-compute/OpenBS.git
claude --plugin-dir ./OpenBS/plugins/openbsOnce installed, invoke any skill with its slash command:
/openbs-review path/to/paper.pdf
/openbs-idea topic:idea machine-learning
/openbs-rebuttal path/to/reviews.txt
/openbs-presentation talk "GPU-accelerated graph processing"
/openbs-thesis qe "distributed systems"
These skills are invoked explicitly via their slash commands (they are marked manual-only, so Claude does not auto-trigger them).
Content is based on Prof. Bingsheng He's resources:
- raintreebook -- A PhD Student's Guide to Navigating the Research Landscape
- CameraReadyChecker -- Automated camera-ready paper validation tool
- His paper corpus (used for the worked examples), on his homepage: Mars (PACT 2008), Relational Joins on Graphics Processors (SIGMOD 2008), Relational Query Coprocessing on Graphics Processors (TODS 2009), Efficient Gather and Scatter Operations on Graphics Processors (SC 2007), Comet (SoCC 2010), and others.
- His group's AI-for-research tools: XtraGPT (ACL 2026) and PaperDebugger (WWW 2026 demo).
Note: this is the systems/database Bingsheng He (何丙胜) at NUS (Xtra Computing Group), not the Nanjing optimization mathematician 何炳生 (ADMM / variational inequalities).
MIT