The thesis is: voice input can widen when coding work fits into a user's day, while review and precise correction remain separate interaction tasks. Willison says he built a newsletters feature for his blog using Codex voice mode in the ChatGPT desktop app against a local development environment. The work included a Django model and migration, import functions, archive pages, and search integration.
The mechanism was a loop across modalities: spoken requirements and clarifying questions drove implementation; a local browser preview made the changing result inspectable; then a GitHub pull request supported review and correction. The mechanism is: spoken direction, local visual feedback, and typed review formed a complementary loop rather than a voice-only workflow. He replaced a Git-subprocess import with an API-based approach and spent another half-hour prompting by keyboard. Why now: the account demonstrates a shipped use of voice coding connected to a real local project, rather than a voice-only demo. Its implication is: voice may make coding agent work compatible with multitasking, but it does not remove the need for review and precise edits.
The evidence boundary is: this is one developer's account of one feature, not a controlled comparison of speed, quality, or general usability. It does not show that the agent independently deployed production-ready software. Watch whether later examples document fewer modality handoffs, or instead continue to show human review and typed corrections as necessary stages. The first watch condition is: look for additional documented projects that report both the voice-driven work and the subsequent review and correction effort.