The flipbook
How a notebook comes together
One conversation, sixteen exchanges, seventeen real commands. You never type them — your agent runs them while you ask questions and direct the work the way you normally would, in a tool like Claude Code. Each exchange shows what was said and what the agent answered; the command it ran sits right under the reply, and "under the hood" expands to the full record — the files that changed, the output, the page that resulted. The story starts with somebody else's evidence and ends with a claim that earned the right to call itself verified — including the moment the human asks for a shortcut and the system tells the agent no.
Every exchange on this page was produced by running the real flip CLI against a scratch notebook when the site was built — the same commands an agent runs — no screenshots, no hand-written output. The research subject is invented: a fictional baking club and a kitchen log, chosen so that no exchange could be mistaken for a real finding.
What the sixteen exchanges are meant to show
Not that flip is easy to use — that the rules are enforced, on the agent, no matter what the human asks for. The exchange worth pausing on is the refusal: the human says "good enough — mark it confirmed," the agent tries, and the CLI declines with an exit code, because the profile says a notebook headed for publication needs two independent judged sources or a grade-A primary. Nothing about that is advisory, and no amount of agent confidence overrides it.
The second thing worth noticing is how little of it is flip's doing. What you are left with at the end of the conversation is a directory of markdown files that any editor can open, plus two append-only ledgers. The discipline lives in the format; the CLI is just the thing that keeps you honest while you write it.
Once the conversation is over, this same notebook is whole — see it rendered end to end, sources through log tail, on the notebook page.