Document Reader

Before first use:

• Read references/install.md
• Confirm the official upstream skills pdf, docx, xlsx, and pptx are available in Claude

If the official skills are installed, prefer their workflow details and scripts as the primary implementation guide.

Directory Structure

This skill family is organized as:

skills/
  document-reader/
    SKILL.md
    references/
      install.md
      output-contract.md
  read-pdf/
    SKILL.md
  read-docx/
    SKILL.md
  read-xlsx/
    SKILL.md
  read-pptx/
    SKILL.md

Routing Rules

Map files by extension:

• .pdf -> /read-pdf
• .docx -> /read-docx
• .xlsx, .xlsm, .csv, .tsv -> /read-xlsx
• .pptx -> /read-pptx

If the user passes a directory:

1. Find all supported files under that directory
2. Group them by type
3. Process each file with the matching wrapper
4. Produce a combined synthesis at the end

If the directory contains no supported files, say so clearly and stop.

Standard Output Contract

For every file processed, follow references/output-contract.md.

At minimum, produce:

• File identity: path, format, and a one-line description
• Structure: pages, sections, sheets, or slides
• Key extracted content: the most important text, tables, formulas, or notes
• Research takeaways: what matters for ARIS-style research work
• Extraction risk: OCR issues, missing notes, broken tables, tracked changes, hidden sheets, or conversion loss
• Next action: what to inspect or ask for next

Workflow

Step 1: Resolve Inputs

• Accept a single file, multiple files, or a directory
• Expand relative paths
• Ignore unsupported file types

Step 2: Infer User Intent

Infer one of these modes from the prompt:

• summary -> concise overview
• extract -> detailed text/data extraction
• structure -> outline, sections, sheets, slides, comments, notes
• compare -> compare multiple files

If the user does not specify, default to summary.

Step 3: Dispatch

• For each file, invoke the matching read-* wrapper
• Prefer Anthropic's upstream official document skill behavior when available
• Keep file-type-specific logic out of this orchestration skill

Step 4: Synthesize

If multiple files were processed, add:

• cross-file overlaps or contradictions
• which file appears authoritative
• what should be cited, quoted, or checked manually

Step 5: Save Notes When Helpful

When the user is processing a research packet or supplementary materials, offer to save a structured summary under:

document-reading/<basename>/SUMMARY.md

Do not create files unless the user asks, or the surrounding workflow clearly expects saved artifacts.

Notes

• Prefer reading the original file rather than converting it unless conversion is required
• Be explicit when extraction quality is uncertain
• For mixed research artifacts, focus on claims, methods, results, limitations, and reproducibility details