In the field of AI office work, many have experienced similar embarrassments: asking AI to generate a report, only for the formulas to turn into static numbers; asking it to modify a PPT template, only for the layout to become a mess. To address this long-standing industry problem of "unusable generated content," the large model manufacturer MiniMax (Xiyu Technology) has recently officially open-sourced its production-grade office document engine Office Skills. This code based on the MIT license aims to make AI-generated Word, Excel, PPT, and PDF documents truly meet the standard of "direct delivery."
Underlying Reconstruction: Bypassing "Average" Libraries, Pursuing High-Fidelity Control
MiniMax's technical choices are quite solid. To solve the issue of "silent data loss" in traditional Python libraries when handling complex formats, they abandoned common but limited solutions:
Excel (xlsx): They gave up openpyxl and instead performed "surgical-level" operations directly at the XML level. By unzipping, modifying specific nodes, and re-packaging, they ensure that existing pivot tables, VBA macros, and complex formulas remain intact.
Word (docx): They abandoned the lightweight python-docx and instead adopted the .NET OpenXML SDK maintained by Microsoft, achieving perfect support for nested tables, headers, footers, and revision tracking.
PDF and PPT: They adopted a "dual-engine" splitting strategy. For example, the cover of PDF is rendered using HTML+CSS to ensure design quality, while the main content is handled by ReportLab for stability; PPT presets four visual "formulas" to ensure global style consistency.
Self-Loop Evolution: A "Digital Employee" with Self-Correction Ability
Aside from open-sourcing the toolchain, MiniMax also shared a self-evolution mechanism called Execute → Evaluate → Fix. This means that this set of Skills is not just static code—it has an automatic evaluation framework. When AI fails to generate a document, the system automatically detects errors in structure, formulas, or formatting, and stores them as repair cases for iteration.
This mechanism makes Office Skills more stable when handling real office scenarios. Whether it's a high-end omakase Japanese menu or a financial report at an investment bank level, AI no longer just provides a "demo," but delivers a final product with a logical loop and compliant formatting. With the open-source of this capability, the office agent race may bid farewell to the "amateur team" era and officially enter the professional delivery stage.
