← Starbound dev log

2026-10-27

Three writing stacks in one night

writing-with-aipublishingai-agents

Flat editorial illustration of three desk setups in a row on a dark background, one desk glowing amber while the other two sit dim

Yesterday I tore down and rebuilt the writing setup for the new [Starbound](https://trystarbound.com) book three times. Not because of some grand plan, but because each tool gave up one specific thing I couldn't work around. If you write with an AI in the loop, the stack matters more than you'd think, and the winning argument surprised me.

First try: Obsidian, free, with a word-count plugin and a pre-built vault. It lasted about twenty minutes. Fine app, but my agent talks to files in git, and I was going to spend my nights shuttling chapters between what the human reads and what the machine edits.

Second try: the polished-markdown route, Ulysses plus Vellum for book production. Genuinely nice tools. But the editing loop means the agent proposes revisions and I accept or reject them, and a proprietary library in the middle of that loop means every pass goes through an export step. For a book that gets revised ten-plus times, that tax compounds.

What actually won, and why

The winner was Google Docs, which I did not expect. Not for the app itself, for the connector. The agent edits the same document I read, in place, and an edit that's accepted keeps its link to the source conversation. The revision history lives where the text lives. One comparison script confirms the docs match the markdown source of truth, so nothing drifts.

And the writing itself is moving. Chapters 1 through 4 drafted yesterday morning, about ten thousand words. Chapters 5 through 8 by night, another eleven-thousand-eight-hundred, plus a continuity ledger, cast voice notes, and a small script that checks voice consistency across chapters. A fresh-context review pass caught five continuity problems in the new batch (a kill tally that didn't add up, a flask count, a numb hand, a bridge, a memory that contradicted an earlier scene). All five fixed before the batch was called done.

The thing I'd pass along: pick the stack by how your collaborator reads and writes, not by which app has the best prose mode. My collaborator is an agent. The document store had to be something it could touch directly. Everything else, I can adapt to.

this is the build log of starbound · trystarbound.com · all entries · essays