Just Talk to It – The No-Bs Way of Agentic Engineering
2 days ago
- #AI-coding
- #GPT-5-Codex
- #workflow-optimization
- The author has transitioned to using GPT-5-Codex for coding, finding it more efficient and reliable than Claude Code.
- They manage a large TypeScript React app, Chrome extension, CLI, Tauri client, and Expo mobile app, hosted on Vercel.
- The workflow involves running multiple Codex instances in parallel within a terminal grid, with agents handling atomic Git commits.
- The author emphasizes the importance of 'blast radius'—understanding the impact of changes before implementing them.
- GPT-5-Codex is preferred for its larger context window (~230k tokens), efficient token usage, and faster performance.
- The author critiques Claude Code for its verbose language, slower speed, and inefficiencies compared to Codex.
- They avoid using third-party harnesses like amp or Factory, finding them unnecessary given the capabilities of Codex.
- The author refactors code regularly using AI, focusing on reducing duplication, dead code, and improving documentation.
- They advocate for iterative development over spec-driven approaches, often starting with minimal specifications and refining features in real-time.
- The author shares tips for effective AI use, such as queueing messages for long tasks and writing tests immediately after feature implementation.
- They maintain an extensive Agents.md file with project-specific guidelines, which GPT-5-Codex adheres to better than Claude.
- Despite its flaws (e.g., occasional panics or lost lines), GPT-5-Codex is praised for its reliability and efficiency.
- The author concludes that managing AI agents requires skills similar to managing human engineers, emphasizing intuition and experience.