20,000 new lines of documentation per month. 500 pages maintained. Five releases per week across nine programming languages. One person.
That's Devin Logan's output as the solo technical writer at Fern, an API documentation and SDK generation tool. The numbers sound impossible until you look at how the work is actually divided between human and AI.
Logan's setup uses five AI tools, but none of them write documentation autonomously. The pattern is consistent: AI handles the mechanical work (file moves, redirects, initial drafts, boilerplate), while Logan handles what AI still can't do well - information architecture, editorial judgment, and recognizing patterns across customer questions.
The Actual Tools and How They're Used
The core of the workflow is a custom Slack agent called "Fern Writer" built on Cognition's Devin, which has full codebase context. Logan triggers it from Slack, and it creates pull requests with drafted pages and reorganized content. Claude Code handles local iteration on content structure. ChatGPT and Claude serve as workshop tools for draft feedback. Notion AI aggregates context from marketing and non-repo content.
A concrete example: when customers were confused by HTTP vs. SDK code snippets scattered across Fern's docs, Logan identified the pattern from support threads, then had Fern Writer update all relevant pages in a single PR. The diagnosis was human. The bulk editing was AI.
Another example shows where AI falls short. When documenting an authentication system, the AI-generated structure incorrectly treated RBAC and API key injection as top-level options. An engineer caught the error - they're actually features within JWT/OAuth flows. The entire information architecture had to be restructured. AI handled the file moves and redirects. The engineer provided corrections. Logan managed the overall structure.
As Logan puts it: "AI is good at a lot of content-related tasks. However, it's not always great at knowing when a paragraph should be three sentences instead of ten, or that documenting a new SDK feature means updating three other pages."
This is a useful reality check for anyone thinking about AI-assisted content production. The productivity gains are real, but they come from offloading tedious mechanical tasks, not from AI doing the actual thinking about what documentation should say or how it should be organized.