15 lessons learned building ChatGPT Apps | OpenAI Developers

This post contributes lessons learned from building two dozen ChatGPT Apps over three months. Key technical insights include managing context asymmetry between user, UI, and model by explicitly defining data visibility; reversing the traditional web development pattern of lazy-loading to front-load data for AI apps; enabling model visibility into UI state via `setWidgetState` and `data-llm` attributes with a Vite plugin; differentiating interaction paths (widget, server, model); adapting UI for inline, PiP, and fullscreen display modes; ensuring UI consistency with the OpenAI Apps SDK UI Kit; implementing language-first filtering by providing models with Lists of Values (LOVs) for tool parameters; and unlocking richer interactions by treating files as first-class inputs for both model reasoning and UI workflows. Production considerations include understanding CSPs as the new CORS, implementing secure MCP tunnels for private servers, and developing robust tooling like the Skybridge framework and a Codex Skill for app development.