📐 Chening

August 14, 2026 · View on GitHub

晨应 · The Chinese National Mathematical Modeling Contest (CUMCM) is 72 hours, one problem set, and 200,000 students fighting for the same medals. I lost every possible way over three years — so you don't have to.

This is an AI Agent skill pack: a decision manual your agent reads to help you pick the right problem, choose the right model, write a defensible paper, and dodge the traps that quietly kill hundreds of teams every September.

Built by a student at Northwestern Polytechnical University (西工大) who spent three Septembers in this arena.

CUMCM Skills License

Why this exists

I played CUMCM three times and stepped in basically every trap there is:

  • Wrote an abstract with no numbers in it — the judges skim three minutes, my result was unreadable
  • Put a figure caption in the wrong place — the judge's eyes skipped right past it
  • Picked problem C, then discovered the data was nowhere near enough

Every single one was a lesson paid for in contest hours.

Meanwhile GitHub is full of AI skill packs — all in English, all aimed at Silicon Valley. None for the hundreds of thousands of students who sit this Chinese contest every year. So I condensed three years of notes into this pack and gave it to your agent, so it can judge your paper like a CUMCM referee, pick your model like a coach, and nag you about sensitivity analysis before you forget.

What's inside

skills/
├── 历年真题速查.md      All 18 problems from 2020–2025, 6 cross-year patterns
├── 选题评估.md          Problem-feature matrix + 6-criterion scoring, 2 worked cases
├── 模型选择决策树.md    Six model families, scene table + flowchart + stats
├── 论文写作规范.md      Abstract 6-elements, 8-section structure, figures, page budget
├── 易错清单.md          25+ traps across 4 phases, fatality-ranked
├── 时间分配.md          72-hour breakdown, hard deadlines, panic plans
├── 图表规范.md          matplotlib palette, three-line LaTeX tables, chart choice
└── 赛前准备清单.md      Hardware/software/templates/team roles, checkable the night before

The skill files are in Chinese by design — that's the language the contest, the judges, and the competitors live in. The pack is a Chinese-market product with an international-facing README.

How to use it

Point your agent (Claude Code, Hana, Cursor, anything that loads skills) at this directory and say something like "load the 数模 skill pack from the skills/ directory."

Each file is a standalone playbook, grab the one for your moment:

  • Choosing A/B/C on Day 1 → 历年真题速查.md + 选题评估.md — a scoring template and two real walked-through cases
  • "What model do I use?" → 模型选择决策树.md — match your data shape and question type
  • Writing / formatting the paper → 论文写作规范.md + 图表规范.md
  • Self-review before submitting → 易错清单.md — ends with a ready-to-paste audit prompt for your agent
  • Mid-contest time check → 时间分配.md — compare against the 72-hour rhythm
  • The night before → 赛前准备清单.md — check off every box

The three iron rules

  1. An abstract without concrete numbers is an abstract that wasn't written. Judges decide in three minutes.
  2. Sensitivity analysis is required homework, not bonus points. Missing it costs you a tier.
  3. Pick your problem by Day 1 noon, freeze your model by Day 3 3pm. After that, switching means rewriting the paper.

系列项目

陈棱是 ChenLaoshiYF 系列开源项目之一：

License

MIT

20 万参赛者。多一个不亏，少一个血亏。

— packaged by someone who's been there. Let your agent carry the scar tissue for you.