Looks promising for document editing with your agent
5 hours ago
- Paper Office is a suite of Python packages for safely and correctly manipulating Word, PowerPoint, and Excel files, building on python-docx, python-pptx, and OpenPyxl.
- Across five models and 61 tasks, Paper packages with guidance passed 92.5% of trials, compared to 80.7% for upstream packages without skills and 69.5% with Anthropic's comparable Office skills.
- Agents using Paper wrote code to edit Office file internals directly in only 1.6% of runs, versus 78.7% without skills and 50.5% with Anthropic skills, reducing context pollution and corruption risks.
- Office Open XML (OOXML) files are ZIP archives with interrelated parts; editing them requires consistency across multiple elements, which standard libraries often lack.
- Paper Office extends the standard libraries with features like cross-run text search, native tracked changes, numbering restarts, content controls, safe slide cloning, reference-aware table edits, and formula cache management.
- Paper DOCX adds document-wide search, native redlines, comment threads, content controls, fields, bookmarks, and cross-document composition, with preservation of package parts.
- Paper PPTX provides effective formatting inspection, relationship-safe slide cloning, formatting-preserving text replacement, merged-cell table editing, and policy-based slide imports.
- Paper XLSX includes package-preserving saves, reference-aware structural edits, formula cache freshness handling, LibreOffice-backed recalculation, and edit receipts.
- In evaluations, Paper improved success rates in most model-by-format slices, with notable gains in XLSX (87.4% vs 74.7% without skills and 44.2% with Anthropic skills).
- Trajectory analysis showed that upstream runs frequently involved manual OOXML repair, while Paper handled these connections via APIs, reducing incomplete edits that leave dependent parts broken.
- Paper also caught formula errors, kept comment threads intact, and updated validation/formatting ranges automatically, outperforming upstream and Anthropic skill setups.
- The authors conclude that software improvements, not just prompts or tools, are a key lever for boosting agent performance in knowledge work.