# CUMCM Skill Pack An installable AI skill for the **China Undergraduate Mathematical Contest in Modeling**: read the problem, model, verify, write the paper, check, and package. Not a prompt dump — templates, scripts, and golden problems included. [中文](README.md) · [2023C demo](examples/2023C-workflow-demo.md) · [2026A demo](examples/2026A-workflow-demo.md) Worked solver for 2026 Problem A: [ciyuan1234/MCM_2026](https://github.com/ciyuan1234/MCM_2026) ## Install ```bash git clone https://github.com/ciyuan1234/MCM_skills.git cd MCM_skills chmod +x install.sh && ./install.sh ``` Windows: `.\install.ps1` Copies `cumcm/` into Claude Code, Codex, opencode, and Grok (`~/.grok/skills/cumcm`). Then ask: ```text Create a CUMCM workspace, read the problem, classify A–E, and propose a route per question. ``` A-type (PDE / heat–mass) loads `references/22-mechanism-pde.md`. C-type starts from `data_contract.json`. ## What you get | Type | Path | | --- | --- | | Data (C/E) | data contract → regression / optimization → V-REG / V-OPT | | Mechanism (A) | unit dictionary, conservation, grid/time-step checks, V-PDE | | Scheduling (B) | feasible baseline first, then V-OPT | Scaffold pre-creates `1_数据/units.md`, `2_代码/0N/model.md`, and `verify_qN.py`. `verify.py` now checks those artifacts. Full index: [`cumcm/SKILL.md`](cumcm/SKILL.md). Changelog: [`CHANGELOG.md`](CHANGELOG.md).