Claude Now Creates and Edits Spreadsheets, Docs, Slides, and PDFs

Claude can now create and edit spreadsheets, documents, slide decks, and PDFs from your prompts and uploads in Claude.ai and the desktop app. The preview feature, powered by a private computing environment where Claude can write code and run programs, supports analysis, modeling, and cross-format conversions. Enable it in experimental settings, guide Claude via chat, and be mindful of data risks due to internet access.

Key Points

Claude can now generate and edit real Excel, document, slide, and PDF files directly in the app and desktop.
It handles data cleaning, analysis, charting, spreadsheet modeling with formulas, and cross-format conversions.
Available in preview for Max, Team, and Enterprise users; Pro access is coming soon.
Powered by a private computer environment where Claude writes code and runs programs, making it an active collaborator.
Setup requires enabling an experimental feature; users are cautioned about internet access and potential data risk.

Sentiment

Mostly negative/skeptical: strong interest in the capability, but widespread frustration with recent reliability and quality regressions overshadows the announcement. Some curiosity and approval of the sandbox design, yet many demand stability, transparency, and fixes first.

In Agreement

Having a built-in, sandboxed code environment to create real Office files is useful—like Code Interpreter—with ready-to-use spreadsheets, charts, and documents.
Cross-format transformations (e.g., PDF to slides, invoices to structured spreadsheets) can save significant time and reduce tedious work.
It’s valuable as a second opinion alongside GPT-based analysis; multiple users reported parity in data analysis workflows.
Clear allowlisting and a proxied environment provide a measure of safety and control over internet access and package installs.
This moves LLMs closer to practical, agent-like collaboration for routine business tasks (spreadsheets, slide decks, reporting).

Opposed

Reliability and quality have degraded recently (especially coding): instruction-following failures, hallucinations, tool-use being ignored, context loss, lag, and frequent artifact-editing glitches.
Users suspect quantization or throttling to handle load, despite Anthropic saying they don’t intentionally degrade quality; capacity issues persist.
New features feel misplaced when core stability is lacking; calls to prioritize fixes and publish transparent, production-eval benchmarks over time.
The feature disables the older Analysis tool and doesn’t work consistently across clients (e.g., iOS), which is confusing and limiting.
Security and governance concerns around granting internet access and agentic code execution; enterprise IT constraints make local scripting infeasible for some.
Accessibility gaps (e.g., no TTS) and broader skepticism about product direction (pivoting away from engineers toward general consumers).