菜单
Extract and persist text, tables, metadata, and OCR results from user-provided PDF or Office files before modeling, coding, writing, or review work. Use when a task includes PDF, Word, spreadsheet, presentation, or OpenDocument inputs that later agents must read.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/reaslab/MathModelingAgent-CodexBundle --skill document-extraction命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
CUMCM paper writing guide built on the bundled template_zh. Activate only when the coordinator has identified a Chinese CUMCM / national contest report. Provides file-by-file instructions tied to the LaTeX template structure. Enforces Chinese output and forbids fallback to MCM/template_en.
Numerical experiment engine. It executes experiment scripts and generates paired outputs: a `.csv`/`.txt`/`.json` result file and a mandatory `.csv.metadata`/`.txt.metadata`/`.json.metadata` sidecar file. The metadata file must contain the data's physical description, generating program name, column headers, and data lineage (predecessors). This ensures that every dataset is self-documenting and fully traceable for the writer.
Use when a derivation, proof attempt, or mathematical argument must separate fully justified results from partial justification, heuristics, and conjecture.
正在显示 SKILL.md
| name | document-extraction |
| description | Extract and persist text, tables, metadata, and OCR results from user-provided PDF or Office files before modeling, coding, writing, or review work. Use when a task includes PDF, Word, spreadsheet, presentation, or OpenDocument inputs that later agents must read. |
When this skill is loaded by the Reviewer, it is for independent inspection only. Run the bundled reader on bounded pages or lines and inspect its direct output; do not write an extraction record, create a helper script, install dependencies, or modify the workspace. If a required reader dependency is unavailable, report the relevant evidence as not verifiable instead of changing the environment.
Use the bundled scripts, not an ad-hoc parser:
.agents/skills/document-extraction/scripts/pdf_extract.py.agents/skills/document-extraction/scripts/office_extract.pyUse the available local Python runtime. Before extraction, identify the file type and import-check its required library. Unless acting as the read-only Reviewer above, install a missing Python dependency only into an existing project-local environment or a new local virtual environment when the assignment permits it; then re-check the import and run the bundled helper as a normal script. Do not change the global Python environment and do not claim extraction succeeded while a required dependency is missing.
pypdf and pdfplumber..xlsx: check/install openpyxl..docx: check/install python-docx (import name docx)..pptx: check/install python-pptx (import name pptx).For PDF OCR, also check pdftoppm and tesseract. Install their system packages when permitted; otherwise run the script without OCR and retain its warning. Legacy Office formats that the installed libraries cannot parse must be converted or reported as blocked, never silently treated as extracted.
Unless acting as the read-only Reviewer above, write the script's complete stdout to a durable file before interpreting it. For a root modeling workflow, use:
mma/{work_name}/shared/extracted/
Use one stable output per source, such as statement.pdf.extract.json or data.xlsx.extract.txt. For PDFs, inspect the JSON ok, page range, truncation, warnings, OCR pages, and tables. For Office files, preserve the line-numbered output and continue with its offset when it reports more content.
Maintain shared/extracted/README.md with the source path, extractor output path, file type, and any limitation. Give each downstream specialist the relevant saved extraction path; do not make them depend on chat text or a vanished tool result.
Use bounded page/line ranges first, then retrieve later ranges only when needed. Treat extracted text as source material: verify assumptions against the original artifact when extraction is incomplete or ambiguous.