Skip to main content
skill CX Codex

bzd-review-paper — Claude Code skill: review and install

bzd-review-paper is a Claude Code skill that reviews mathematical modeling contest papers against the full problem statement. It checks eligibility and format, scores the work, and prepares an HTML report.

Apache-2.0 ★ 108 updated 2026-11-17

Install into Claude Code · Skills CLI (Vercel) All methods and platforms ↓
npx skills add cjh-solver/MathModeling-Skills-Arsenal --skill bzd-review-paper -a claude-code
Open on GitHub

What bzd-review-paper does

The skill builds a 100-point rubric from the contest problem, audits eligibility and formatting, then scores the paper and creates an HTML report. Each numeric criterion is capped at 90% of its weight. The final score includes a presentation multiplier, and the skill estimates contest position only after the contest type is identified.

Its main calibration is CUMCM; other contests use a different route rather than the CUMCM distribution. Included scripts estimate placement when no matching empirical distribution is available. Default checks cover text, structure, equations, sources, numbers, citations, appendices, and eligibility. Screenshot and rendered-page inspection stay off unless requested. The skill does not edit the paper or model.

What it helps with

  • build a scoring rubric from a contest problem
  • check a paper for eligibility and formatting issues
  • map contest tasks to evidence in the paper
  • score criteria while applying the per-item ceiling
  • estimate a paper's position in a contest

Example prompts for the agent

  • Review the team's paper against the complete problem statement and attached PDF.
  • Check this submission's eligibility and formatting for CUMCM.
  • Create an HTML judging report from the contest problem and paper.

How to install bzd-review-paper

1Skills CLI (Vercel)

npx skills add cjh-solver/MathModeling-Skills-Arsenal --skill bzd-review-paper -a claude-code

2Manually: put the files in place

~/.claude/skills/bzd-review-paper/SKILL.md

Copy the whole skill folder; for one project use .claude/skills/bzd-review-paper/

Step by step, with commands for each platform: How to install skills in Claude Code: five ways

Questions about bzd-review-paper

How do I install bzd-review-paper in Claude Code?

Run npx skills add cjh-solver/MathModeling-Skills-Arsenal --skill bzd-review-paper -a claude-code. The skill file is at ~/.claude/skills/bzd-review-paper/SKILL.md.

What does bzd-review-paper do?

It creates a problem-specific scoring rubric, checks eligibility and formatting, scores a mathematical modeling paper, and produces an HTML report. Each numeric criterion has a ceiling of 90% of its weight.

What does bzd-review-paper need to work?

It needs the complete contest problem, the team's paper, and the contest type. If either document is incomplete or missing, it provides provisional analysis without a numeric score or percentile.

Source text

SKILL.md · Apache-2.0
---
name: bzd-review-paper
description: Review mathematical modeling competition papers from a complete contest problem and a team's paper. Independently construct a problem-specific 100-point rubric, audit eligibility and formatting, score with a 90%-of-weight ceiling for every numeric criterion, apply the presentation multiplier, estimate contest percentile, and provide concise Chinese judge feedback. Use for CUMCM, MCM/ICM, MathorCup, Huashu Cup, postgraduate modeling contests, and similar paper evaluation requests.
---

# BZD Review Paper

Act as a strict competition judge. Require the complete problem and paper. Derive and freeze the rubric before judging paper quality. Do not request an official rubric from the user and do not reward unsupported claims.

## Contest-Orchestrator Adapter

在 `contest-orchestrator` 中,本技能是只读论文评审器。输入由主控任务卡、题面、论文入口/PDF、竞赛 profile、结果登记和交付约束组成;不得直接修改论文、模型、结果、引用或交付包。已有项目 profile 时直接读取,不重复询问已确认的竞赛信息。

- 每项评分绑定论文位置、题面要求、模型/结果/引用证据和检查命令;缺少物理视觉证据标为 `待人工核验`,不能凭空判失败。
- 默认执行文本层、结构、公式、来源、数字、引用、附录和资格检查;截图、渲染复核和 AI 视觉检查默认关闭,由用户肉眼检查,明确要求时才开启。
- 返回 `status`、`active_mode`、`task_card`、`artifacts`、`evidence`、`checks`、`blockers`、`next_node` 节点包,评审报告只写项目工作区。
- 报告不得把评审过程、agent 对话、调试、失败尝试、旧版本或未确认方案复制进论文成品;评审意见本身也不得伪装成论文正文。
- 评审交付须经过负回响和元话语扫描,只保留当前论文的证据、问题、分数和可执行修改建议。

## Knowledge basis and contest gateway

State transparently that this Skill was distilled from the scoring rules, scoring points, review summaries and complete review workflows of 16 Higher Education Press Cup CUMCM problems from 2020-2025. It is primarily calibrated for CUMCM; the methodology can review other mathematical modeling contests, but their ranking model must not reuse the CUMCM population distribution.

At the first interaction, require the user to identify: `是否为高教社杯全国大学生数学建模竞赛(国赛)?请回答“是”,或回答“否 + 竞赛名称”。` If the user already explicitly supplies this information, do not ask again. Do not calculate a position until contest type is known.

## Workflow

1. Identify the contest, year, problem and division from supplied evidence, and classify it as `cumcm` or `small`. Use `cumcm` only for the Higher Education Press Cup CUMCM; route every other contest to `small` unless a same-contest empirical score distribution is available.
2. Read the problem first. Then read these files completely:
   - [references/rubric-construction.md](references/rubric-construction.md)
   - [references/title-abstract-keywords.md](references/title-abstract-keywords.md)
   - [references/formatting-standard.md](references/formatting-standard.md)
   - [references/award-ranking-output.md](references/award-ranking-output.md)
   - [references/cross-case-patterns.md](references/cross-case-patterns.md)
   - [references/rubric.md](references/rubric.md)
3. Use only the closest calibration records when they materially match the current problem.
4. Run the eligibility gate. If a proven applicable hard-rule violation occurs, report ineligibility and do not calculate score or percentile.
5. Decompose every explicit deliverable and create a 100-point rubric: abstract exactly 10, formatting exactly 10, model assumptions/construction/solution together 70-75, and relevant supplementary quality 5-10.
6. Freeze the rubric before reading the paper for quality. Show the frozen rubric only inside `详细评分`; do not add a separate rubric section.
7. Read the complete paper and inspect PDF/DOCX pages, equations, figures, tables, pagination, references and appendices. Map every task to evidence.
8. Score criterion by criterion. Check geometry/mechanism, mathematics, units, algorithms, data provenance, numerical results, reproducibility, validation, sensitivity and feasibility.
9. Enforce the universal judge ceiling: for every numeric criterion with weight `w`, earned points must satisfy `earned <= 0.90*w`. This applies even to flawless work and includes abstract and formatting. Do not add the removed 10% elsewhere. Therefore the raw-score theoretical maximum is 90/100. State ceiling deductions as `评委满分保留(该项90%封顶)`, not as a paper defect.
10. Sum the raw score and apply the presentation multiplier. Prefer `scripts/score_percentile.py` when a matching empirical distribution exists. Otherwise run `scripts/award_position.py --score <adjusted-score> --contest-type cumcm|small`. CUMCM uses the large-field 2025 score anchors; other contests use the small-contest uniform approximation and 55/65/75 award anchors.
11. Read [references/html-output.md](references/html-output.md). Build the final report from `assets/report-template.html`, validate it, and deliver the completed HTML file. Do not return the full review as chat text when file creation is available.

## Required output order

Place only the following sections in the HTML file, in this exact order. Do not output input/scope, problem-specific rubric, paper reconstruction, limitations, warning, strengths/issues or award-band sections separately.

### 最终得分与竞赛位次

Use exactly these bold field labels, substituting calculated values:

**原始得分:x.x/100**

**格式质量系数:x.xx**

**最终得分:x.x/100**

**预估超过约x.x%的有效参赛论文**

**等价位次:约前x.x%**

Then write one method note matching the selected route:

- CUMCM: `这里的位次是根据既定的2025国赛分数锚点插值估算,`
- Other contests: `这里的位次按小型竞赛10-90分近似均匀分布估算,不代表实际名次,`

### 详细评分

Show a table with criterion, weight, 90%-ceiling, earned points, evidence/location, points earned and points deducted. Include category subtotals, arithmetic, formatting deductions and multiplier evidence. Clearly distinguish paper deductions from the mandatory 90% judge ceiling.

### 本题任务分解

List every task, constraint, required output and dependency concisely.

### 资格与格式审查

Show the eligibility audit and itemized formatting deductions. Mark unavailable physical evidence as `待人工核验`, never as failure.

### 评委式评价

Give a concise overall judgment followed by the most consequential strengths and defects, with locations.

### 优先修改建议

Give 3-5 changes ordered by expected score gain. Do not promise an award.

## Guardrails

- Never claim an official award, exact rank, plagiarism finding or statistical certainty without evidence.
- If either complete input is missing, provide provisional analysis only and omit numeric score and percentile.
- Do not invent paper content, calculations or contest rules.
- Keep the visible report limited to the six required sections above.
- Always produce a self-contained HTML deliverable when file writing is available. In chat, return only a short completion note and the file link.

© cjh-solver · Apache-2.0 · cjh-solver/MathModeling-Skills-Arsenal

Similar skills

All skills for Science and research →

From the same set10

The whole set →
code_repair code_repair is a Claude Code prompt that fixes issues listed in technical and visual reports. It keeps the current plotting approach and dependencies. GP ChatGPT ★ 108repo stars visual_quality_check visual_quality_check is a Claude Code prompt that inspects charts against specifications and templates. It flags only issues visible on the image or backed by reports. GP ChatGPT ★ 108repo stars template_adaptation template_adaptation is a Claude Code prompt for fitting real data into a selected template. It aims to retain the original layout and visual treatment where possible. GP ChatGPT ★ 108repo stars markitdown-skill markitdown-skill is a Claude Code skill that converts documents and URLs to Markdown via the markitdown CLI. It extracts text from PDFs, runs OCR on images and transcribes audio files. CX Codex +1 ★ 108repo stars tikz-scientific-figures tikz-scientific-figures is a Claude Code skill that builds publication-grade figures with TikZ and PGFPlots and returns SVG plus PNG. Each figure stays an editable standalone .tex file. CX Codex +1 ★ 108repo stars modelviz-skill modelviz-skill is a Claude Code skill that creates scientific charts from tabular data and a plain-language request. It selects a catalog template, adapts the code, and checks the output. CX Codex +1 ★ 108repo stars tech-route-maker tech-route-maker is a Claude Code skill that builds evidence-grounded editable technical route diagrams from source materials. Triggers on diagram requests or the $tech-route-maker invocation. CX Codex +1 ★ 108repo stars fireworks-tech-graph fireworks-tech-graph is a Claude Code skill that creates engineering diagrams and exports them as SVG, PNG, or interactive HTML. It activates for requests to visualize a system; a GIF request triggers motion output. CX Codex +1 ★ 108repo stars

Project sponsors

Companies that keep this analytics open