I produce everything by speaking — these pages included
WORKFLOWRepeatable working flow.VOICEVoice-driven production.ACTIVE-USECurrently in use.
I produce this site without typing a single line. I record my own voice conversations, tag the useful ones in the panel, then have agents produce from those conversations. The marketing copy, these guide pages, the what’s-new entries — all of it comes out of this loop.
The visual below is the very conversation that asked for this page: guide-tagged raw speeches sit in the panel with the production prompt placed on top as it is. The agent receives the speeches and the prompt as one packet.
Two nested visuals: the final step takes you to the complete panel — list, detail, audio and search in one place.
The whole cost is the up-front speeches. The raw conversations behind this page took 3, 5, 2, 3 minutes and 18 seconds. The total cost is that speaking time spent up front; from there it proceeds by resolving on the agent side. The panel counter showed 49 hours 37 minutes of accumulated speech while these lines were written — the whole system was built with those conversations.
The task reaches the agent by speech too. Below is a Ghostty terminal: the entire input is a speech transcript, not a single hand-typed line inside. The transcript is pasted with its markup; the agent resolves rectangles, images and timing from that text.
What if the transcript comes out wrong? That is the first question people get stuck on. The answer is simple: in a 20-minute conversation, even with typos — even if I say something wrong — it causes no problem, because I can recover it later. The rest of the speech corrects the mistake; the model separates the confusing part from the whole.
There is a small trick too. I add this sentence to the end of the speech: “I don’t know if I should do it this way — find whatever is right.” My thinking shifts mid-speech — first I say A, then B, then C. With this sentence I hand the decision to the model: I collect my feelings and the options, the model picks what is right.
The same speech packet can be given to more than one agent. This page was written in a Claude Fable 5 session; the same packet was handed, in the same form, to a Codex agent running GPT 5.6 Sol xhigh. The packet is agent-independent: intent, evidence and references travel inside the conversation, so which agent produced it stays an implementation detail.
The tool side of this loop is told on its own pages; the cards below are the interactive visuals of the same release.