What the formatting research was designed to answer
A useful ChatGPT to Word workflow must do more than copy visible text. It needs to remove chat-interface presentation styles while retaining the semantic structures a person expects to edit in Word.
The checks therefore compare normalized text order, document nodes, generated OOXML, coverage warnings, and the final browser download.
Does the conversion keep the same visible words, punctuation, numbers, and link text?
Do headings, lists, tables, links, and code become real editable Word objects?
Are source fonts, sizes, backgrounds, classes, inline styles, and copy noise excluded?
Does unsupported content create a stable warning instead of disappearing silently?
Tested claims and current status
The current claims come from rich HTML, Markdown, plain-text, cross-model-style fixtures, OOXML assertions, browser downloads, and a Microsoft Word open check. They describe the supported first version rather than every possible chat layout.
| Area | Status | What the evidence shows |
|---|---|---|
| Visible wording and order | Validated | Text remains unchanged apart from NBSP normalization and removal of zero-width copy characters. |
| Headings, emphasis, and links | Validated | Generated OOXML contains real Heading styles, text runs, and external hyperlink relationships. |
| Two-level lists and simple tables | Validated | Numbering definitions, editable table cells, and tested table geometry are present in the DOCX. |
| Code blocks and inline code | Validated | Code text stays editable and uses controlled monospace Word styling. |
| Source presentation styling | Removed | Fixture backgrounds, arbitrary fonts, sizes, classes, inline styles, and unsafe nodes do not enter the output. |
| Images, complex formulas, nested tables | Warning | Unsupported objects create explicit coverage warnings and are not counted as fully preserved. |
Input and resource boundaries
The tool accepts one pasted response and automatically chooses safe rich HTML, Markdown, or plain text. It does not depend on private ChatGPT, Claude, or Gemini class names.
| Area | Status | What the evidence shows |
|---|---|---|
| Visible characters | Up to 200,000 | Larger input stops before expensive processing and asks the user to shorten it. |
| Clipboard HTML and DOM | 2 MiB / 10,000 nodes | Hard limits protect the browser from oversized or unusually complex copied markup. |
| Simple table | Up to 100 × 20 | Larger or nested tables are flattened or stopped with a visible warning. |
| Processing window | 10 seconds | The task stops before later export stages when the deadline is exceeded. |
Privacy and claim boundaries
Pasted text, HTML, links, the normalized document, and the generated DOCX stay in the browser. Analytics use only fixed categories and broad count ranges.
The converter does not rewrite, humanize, detect authorship, remove an AI watermark, or promise detector bypass. ChatGPT is used only to describe a compatible workflow; DocReviewKit is not affiliated with or endorsed by OpenAI.