I code daily using Codex App, Antigravity, and @trae, switching between them for different tasks. I've built 4 apps and use all three for specific sets of tasks. Here are my notes on each:
In @antigravity:
Gemini 3 Pro and Flash inside Antigravity excel at UI tasks and flow testing. I work with TDD logic, and each session is lint-free-I keep 0 lint errors and warnings. I don't let them stack and ensure almost no "any" types remain after a few sessions. UI tasks execute almost perfectly, including complex animations and elements. Gemini also handles translations, i18n tasks, and manual improvements in translation files very well. It manages agentic functionality implementations great in TDD mode. I don't give Gemini auth, DB, or payment-related tasks-those go to Opus 4.5, Codex 5.3 Extra-high, and Codex 5.2 (for architecture research and optimization planning).
Codex 5.3 in the new Codex App currently controls Supabase, commits, and deployment-and handles them well.
@trae_ai with Codex 5.2 in solo mode runs multiple planning and research sessions with different agents, tracking vulnerabilities and architectural gaps. I use individual system prompts for each agent in Trae, including a core orchestrator prompt to spawn relevant agents.
In Antigravity and Codex App, I don't use system prompts. In Antigravity, I run multiple planning session loops until the plan fully satisfies me and includes all relevant automated flow testing. Each planning loop uses a detailed, targeted prompt covering related flows and functionalities. I also "cheat" by using @OfficialLoganK's Google AI Studio to build simple functions and UI pages, then merge them into the code with Antigravity-using Opus 4.5 for function-side work and Gemini 3 (in both Trae and Antigravity) for UI/UX parts.
In the new Codex App, I make very targeted changes after a single-cycle implementation planning session for several specific tasks. I especially love using function calling with it-it performs extremely well.
Overall, I don't automate large chunks or long sessions. I prefer frequent manual testing. I'm experimenting with multiple models, but this is currently my most stable workflow. I'm not rushing to 4.6 (though I've tried it). Codex handles the same tasks at a similar level, so I'm not hyped about it yet. I've been working great with 4.5, and the recent changes mostly benefit those chasing full autonomy for long-running tasks. I avoid hype distractions and focus on using existing systems effectively at optimized cost.
Drop your system in the comments and some of you will receive a free annual subscription for 2 of my products I'm launching soon-share your coding setup!
P.S. I write posts manually but with @grok enhancements. I use Veo 3.1 Fast to generate long/short format videos, Nano Banana Pro for image editing and complex infographics. Q, based on Gemini Live's agent, performs all tool calls and real-time actions across multiple of my apps and performs amazingly. Gemini models perform better in overall tasks... but that's IMHO.
English
