source-document-extractionlisted
Install: claude install-skill vincentdania/agent-skills-for-real-work
# Source Document Extraction
Use at the start of any task that depends on the contents of a document. Extraction errors propagate silently into every downstream conclusion, so this step is worth doing carefully rather than quickly.
## Decide the document type first
| Type | Test | Approach |
|---|---|---|
| Digital text PDF | Text selectable, copy works | Direct text extraction |
| Scanned PDF | Text not selectable, page image only | OCR, with page-level accuracy check |
| Mixed PDF | Some pages selectable, some not | Per-page detection, then route each page |
| DOCX | Native Word file | Structured extraction preserving headings, tables, lists |
| XLSX/CSV | Spreadsheet | Read as data, not text; preserve sheet and column structure |
| Email | Message file or mailbox | Extract headers, body, and attachment inventory as separate items |
Never assume a PDF has a text layer. Test it.
## Workflow
1. **Inventory the inputs.** List every file with its type, page count or row count, and size. Confirm you have the complete set before extracting. Reviewing five of six annexes produces a wrong verdict.
2. **Detect the type per document, and per page where mixed.**
3. **Extract preserving structure.** Headings, numbered lists, and tables carry meaning. Flattened text loses which figure belongs to which column.
4. **Recover tables as data, not prose.** Read tables into a structured form and check column alignment and header detection. Misaligned tables are the most common silent