Simon Willison Codes Feature Using OpenAI GPT-6 Astra

Developer Simon Willison built and shipped a new Newsletters section for his website almost entirely by talking to his computer. Utilizing the ChatGPT desktop app Codex tab in voice conversation mode, Willison tasked OpenAI's GPT-6 Astra High model with creating a Django feature while he cooked dinner. Over a half-hour hands-free session, the AI agent built database models, configured Django Admin, and created public archive pages including /newsletters/ and /newsletters/2026/.

The model autonomously navigated data fetching for both public and private sources. GPT-6 Astra High called Substack's undocumented /api/v1/archive endpoint, executing a web search to locate an article by Karen Spinner to resolve pagination logic. It also pulled archived monthly updates from Willison's public simonw/monthly-newsletter-archive repository alongside a private GitHub repository for sponsors-only content, tweaking layouts based on vocal feedback against a local preview.

After completing the initial build, Willison instructed Codex to open a GitHub pull request. Reviewing the code revealed that Astra had chosen a Git subprocess for imports; Willison switched to typing for an additional half hour to refactor the script to use direct API calls for private repositories. The final deployment combined reverse-chronological Substack posts with monthly sponsor updates across the site.

For AI practitioners and software engineering teams, the experiment highlights the practical utility of voice-driven AI agents when paired with real-time visual previews. While direct keyboard input remains far superior for precise code refactoring, pasting error logs, and detailed syntax selection, voice interaction provides a surprisingly effective workflow for high-level architectural work, rapid prototyping, and hands-free multitasking outside shared workspaces.