Ben Balter, a developer advocate and author, is making a compelling case that the era of 'garbage in, garbage out' has arrived for AI assistants. In a recent blog post, Balter argues that documentation quality is now the primary bottleneck for effective AI integration in development workflows. The core thesis is simple: if your institutional knowledge is trapped in unstructured Word documents, your AI tools are effectively blind. Teams with scattered tribal knowledge get generic responses, while those with well-structured Markdown get context-aware answers that actually solve problems.

The Cost of Proprietary Formats

Balter illustrates the problem by comparing file sizes and structure. A sample document measured 33,621 bytes in Word format, compared to just 80 bytes in Markdown. While modern context windows are large, the 'Lost in the Middle' phenomenon means models struggle with irrelevant or poorly structured data. Word files are described as 'tiny jails for content' because they bundle text with proprietary XML formatting that obscures semantic structure. When a model encounters a bolded line of text instead of a declared Heading 2, it loses the hierarchical context needed to retrieve specific information accurately.

Practical Tools for the Migration

To bridge the gap, Balter highlights two new tools: word2md.com and md2word.app. Both operate entirely in the browser, ensuring no data is uploaded to third-party servers, which is critical for enterprise security. For developers who prefer the command line, these functions are also available via npm packages (npx word-to-markdown and npx markdown-to-word). Furthermore, both tools are exposed as MCP (Model Context Protocol) servers, allowing AI agents like Claude or Copilot to convert documents on the fly. This integration means your agent can ingest a legacy .docx, convert it to Markdown for analysis, and export a final Word document for legal review without leaving the chat interface.

A Playbook for AI-Ready Docs

The article provides a concrete playbook for teams. For non-technical staff, the directive is to use real heading styles in Word or Google Docs, not just bold text. For developers, the advice is to put all Markdown documentation in version control, treating AGENTS.md and CLAUDE.md as standard repository files. Balter shares a personal anecdote where an LLM auditor caught a factual error in his own 99,000-word book simply because the text was structured in Markdown. The lesson is clear: structure makes review cheap, and cheap review makes quality scalable.

Key Takeaways

  • Use semantic headings (H1, H2, H3) in all documents; visual formatting like bold text is invisible to AI structure parsers.
  • Migrate canonical knowledge to Markdown or properly styled Google Docs, treating Word files as disposable export formats.
  • Leverage MCP servers like word2md to let your AI agents handle format conversion autonomously.
  • Store all documentation in version control to enable diff-based reviews and automated LLM auditing.

The Bottom Line

Stop letting your AI choke on XML. If your docs aren't in Markdown, your AI isn't working for youβ€”it's just guessing.