Zenn writer lets AI agent draft every post via GitHub pull request

The author's starting point was a habit of building something new before checking what already existed, so this time the options were listed first: writing in Zenn's web editor, pushing hand-written posts to a linked GitHub repository, or letting an agent write them. Zenn already offers an official GitHub integration, so the only missing piece was an agreement on what to delegate to the agent and what to keep for a person — and that agreement became a file.

The setup keeps one source of truth: a private repository holding articles, books, and images. A phone is where the author talks with the agent, a PC is where display is checked, and the repository lets any environment start from the same state without moving a conversation between devices. The first commit went straight to main because an empty repository has no branch to open a pull request against.

The rules also set boundaries on content: no information from the author's employer or work, no personally identifying details, no API keys or passwords, and no copying of others' text. Articles must distinguish things actually tried from things merely researched, and product specs must carry the date they were checked. Conversation logs tend to carry personal details, so the agent strips them at the draft stage — but the author kept the final check as human work, on the reasoning that handing publication to the agent would remove that checkpoint.