What it does

Supervisor-Skills is a dual-track knowledge base: a Chinese-language handbook covering research methodology, paper writing, scientific plotting, and case studies from ICML/ICLR/VLDB papers, plus a set of copy-pasteable “AI Skills” (structured prompts) for tools like Claude, DeepSeek, and Kimi. The goal is to make a busy professor’s implicit judgment — what makes an idea worth pursuing, whether your introduction flows, if your figures meet top-conference taste — available on demand.

The interesting bit

The project treats prompt engineering as pedagogy. Rather than generic “write me a paper” prompts, it encodes specific academic heuristics: a 5-dimensional idea-evaluation framework (higher, faster, stronger, cheaper, broader), an Introduction flowchart model, and a pre-submission reviewer that checks against actual writing checklists and common English errors. The “Vibe Research” section also explicitly addresses how to use AI tools without letting them become, as the author puts it, “expensive toys.”

Key highlights

• Seven executable skills: Idea Evaluator, Vibe Research Guide, Introduction Drafter, Tech/Benchmark Paper Templates, Pre-Submission Reviewer, and Figure Design Advisor
• Case-study dissections of three recent accepted papers (Alpha-SQL at ICML 2025, AFlow at ICLR 2025, LEAD at VLDB 2026)
• CC BY-NC-SA 4.0 license; explicitly non-commercial
• Early-stage project with active TODO list including rebuttal writing and academic collaboration guides

Caveats

• Content is primarily in Chinese; English README exists but the depth is unclear
• The “Skills” are essentially sophisticated prompt files — you still need the underlying model and your own data
• Author notes the project is early-stage and invites contributions, so coverage gaps are expected

Verdict

Worth bookmarking if you’re a Chinese-speaking grad student in data systems or ML who wants structured guidance on the unglamorous parts of research — figuring out if your idea is actually novel, whether your plot conveys what you think it does, and catching “obvious” errors before reviewers do. Less useful if you already have an attentive advisor with time to spare, or if you work far outside the covered domains.