SKILL.md

v1.0.0 · 2263 B · text/markdown
---
name: document-data-extraction
description: Use when reading or generating PDF, DOCX, or XLSX files - work from the actual file structure instead of guessing contents, so output reflects what's really in the document
---

# Document Data Extraction

Summarizing or editing a document by pattern-matching its filename or a
partial preview invites hallucinated content. This skill's rule: read the
actual structure before making any claim about what's in a document, and
edit through structure-aware operations, not blind text replacement.

## Reading

- **PDF**: extract actual text content (and note where a page is
  image-only and would need OCR rather than silently skipping it or
  guessing its content).
- **XLSX**: read actual cell values and formulas, not just what a preview
  shows — a formula's computed value and its underlying formula can both
  matter depending on the task.
- **DOCX**: parse the real document structure (headings, tables, styles),
  not just a flattened text dump — structure often carries meaning (a
  table's columns, a heading's level).

## Generating and editing

- **Never fabricate contents to fill a template** — if a required field's
  value isn't available, say so explicitly rather than inventing a
  plausible-looking placeholder that looks like real data.
- **Edit existing files by modifying their structure**, not by regenerating
  the whole document from scratch — a full regeneration risks losing
  formatting, styles, or content outside what was explicitly asked to
  change.
- **For spreadsheets, preserve formulas where the task doesn't require
  changing them** — overwriting a formula cell with its current computed
  value silently breaks the sheet's behavior on future edits.

## Verification

After generating or editing a document, re-read the actual output file
before declaring the task done — confirm the change landed correctly in
the real file, not just that the generation step didn't error.

## What this isn't for

Simple, single-value lookups from a short document don't need heavyweight
structural parsing — reserve the full structural approach for anything
where getting the content wrong would matter (financial data, contract
terms, anything downstream systems will consume).