Skip to content
Call wiki

How do you move words between AI agents and your Docs, Word, and Sheets?

Last updated 2026-08-18 · Overview — no new claims; each statement links to the entry that carries the primary sources.

Agents read and write plain text. Word, Excel, and Google Docs are foreign formats they reach through machinery — so most of what I set up on calls keeps the work in text files as long as possible and crosses into document-land once, deliberately.

AI loves text files

Section titled “AI loves text files”

A markdown or HTML file is text the model edits directly — and the bulk of what it trained on. A .docx is a zipped XML archive the agent manipulates through code, blind to the rendered page: the vendor’s own Office tooling unzips the file, rewrites raw XML, and renders screenshots just to see the result (Markdown vs Office). Spreadsheets follow the same rule: download as CSV and hand over the text file — it works on every surface, while a pasted Sheets link loads nothing and raw .xlsx pays off only where the agent can run code (Feeding spreadsheets).

Draft in markdown; convert once

Section titled “Draft in markdown; convert once”

The habit I teach: keep intermediates and drafts as markdown files in the working folder, not Google Docs — short, modular, dated. Those files are the durable output of a session; the next session doesn’t reference past chats, it reads the files. When layout matters, draft in HTML — the web is what models trained on — and convert the first draft to the deliverable format early, once, with a screenshot check, so conversion surprises don’t land on the deadline (Markdown vs Office). Repeated edits to a Word file that make it worse each round are the tell you’ve left this path.

Writing back to Google Docs

Section titled “Writing back to Google Docs”

Three routes, by ambition (Writing to Docs). New document: have Claude create the file and save it to Drive, formatting intact. Formatting only: enable Markdown in Docs and use Paste from Markdown — it keeps the headings, bold, and bullets that plain paste drops. Editing an existing Doc in place — one link, version history intact — needs the Docs/Sheets APIs, directly or through a CLI wrapper; the chat connector only uploads files, so a re-upload is a new document that abandons the shared link and history. Driving the Docs web page by screenshot-and-click exists too, but none of those guarantees apply there — what I say on calls: don’t rely on the browser route, the raw APIs work.

Comments and track changes

Section titled “Comments and track changes”

The claude.ai connector strips both comments and suggested edits — worse, suggested text can arrive flattened into the prose, so the model critiques crossed-out sentences as if someone wrote them (Comments & suggestions). Google’s APIs expose both: suggestions with exact positions, comments anchored by the passage they quote — and an agent can write its edits as suggestions rather than direct changes. The chat-level workaround is screenshots: vision sees the strikethrough that text extraction drops.

Review without the copy-paste round trip

Section titled “Review without the copy-paste round trip”

Iterating on prose in chat means pasting blocks back and forth and hunting for what changed. Comment-mode editors replace that: Roughdraft opens agent-written markdown in a Docs-style review interface — comment, suggest, hit Done, the agent revises; Lex does it inside its own editor; and since August 2026 agents can leave native comments and suggestions on Google Docs themselves (Comment-mode editors). The move I recommend for a colleague’s Doc: download it to markdown, review locally, sync back.

Go deeper

Section titled “Go deeper”