Building a blog feature by voice while cooking dinner

Simon Willison wanted a newsletters page for his blog, one that indexes his free weekly Substack issues and his monthly sponsors-only briefings. He built most of it by talking to the ChatGPT desktop app's Codex mode while cooking dinner, with the model, GPT-6 Astra High, pointed at a local dev server so he could watch the site change as he spoke. After about half an hour of conversation the feature was nearly done. He then had Codex open a pull request, reviewed it on GitHub, and spent another half hour typing prompts to fix the rest before deploying.

The work itself was not trivial. Codex added a Django model and migration, admin configuration, and four separate import paths: recent Substack posts over RSS, older ones through an undocumented archive API whose pagination it found by searching the web, published monthly newsletters from a public GitHub repo, and the latest private issue from a private repo. It built a public index and per-year archive pages, put monthly issues into site search while leaving out the Substack copies that duplicate existing content, and when Simon read the pull request he caught one import shelling out to git and had it switched to an API call, because that source repo was private.

His takeaway is plain. Voice is good for high-level direction, and the live visual preview made it far more useful than talking to a model on a phone with no screen. Typing is still better for the precise parts, like pasting an error message or pointing at a specific line. He does not expect voice to become his daily driver, but it let him build something real while his hands were busy.

Why it matters

If you write code with an agent, this maps out where voice actually helps: steering a first draft while you do something else, not the close work of debugging. Worth trying for the rough shape of a feature, then switching back to the keyboard for the fixes.