Skip to content

reference-document-review

Pack: course

Read, normalize, compare, extract, or convert reference materials in PDF,

Compatibility: Designed for OpenCode. Assumes repo-local .opencode/skills discovery

Purpose

Turn heterogeneous reference material into structured course-development inputs.

Default intake workflow

  1. Identify each source and its role:
  2. authoritative source
  3. rough draft
  4. supporting example
  5. visual reference
  6. conflicting reference
  7. Extract the useful material:
  8. objectives
  9. concepts
  10. terminology
  11. data points
  12. examples
  13. diagrams/tables
  14. Normalize the extracted content into a course-ready note.
  15. Flag uncertainty, OCR risk, or contradiction.
  16. Route the cleaned result into the next relevant artifact.

Review modes

  • summary
  • structured extraction
  • comparison/diff
  • lesson-input conversion
  • issue spotting
  • terminology normalization

Expected output

When reading references for course work, prefer this structure:

# Source Review

## Source inventory
## Key extracted points
## Reusable course material
## Conflicts / uncertainties
## Suggested downstream artifact

Source handling rules

  • Keep source facts separate from your suggestions.
  • Preserve citations or file attribution when available.
  • If a scan or image is ambiguous, mark it instead of guessing.
  • When multiple references disagree, present the disagreement explicitly.

Format handling

对于非Markdown格式的参考资料（PDF、DOCX、XLSX、HTML、PPTX等），优先使用 doc-reader skill将其转为Markdown后再进行审阅：

uv run <doc-reader-path>/scripts/doc_to_markdown.py input.pdf --output temp/review.md

然后在Markdown基础上执行审阅workflow。

对于PPTX，优先使用 ppt-reader 获取结构化inventory，再用 doc-reader 补充全文（含表格和图表文字）。

当 doc-reader 未安装时，回退到agent原生文件读取能力，但应明确提示转换质量可能不稳定。

Common source-specific defaults

  • PDF: capture title, section structure, tables, and figures.
  • DOC/DOCX: preserve editorial structure and tracked-intent where obvious.
  • Markdown: preserve heading logic and embedded code blocks.
  • Image: extract visible text, diagram meaning, and any labels that influence course design.

Gotchas

  • Do not silently merge conflicting source statements.
  • Do not treat a marketing deck as authoritative technical truth.
  • Do not rewrite source intent too early; first extract, then reinterpret.
  • Do not discard appendices if they contain definitions or lab constraints.

... (2 more lines in full SKILL.md)