Leading Developer Overhauls AI Workflow: Embraces GPT 5.5, Remote Agents, and Conversational Prompting

A prominent developer has revealed a complete overhaul of their AI-driven software development workflow, marking a significant departure from previously favored methods. Initially reliant on tools like Cursor, detailed plan mode, and Claude Opus models, the updated approach now centers on GPT 5.5 for core model inference, citing “effectively unlimited inference” on their plan. The transition also prioritizes the Codeex harness for its robust CLI and app capabilities, complemented by T3 Code, a self-developed open-source application designed to manage multiple AI harnesses and optimize agentic development. A critical aspect of this evolution is the embrace of remote development environments, leveraging setups like Mac Mini, Helium, and even Replit for hosting T3 Code instances, addressing the challenge of continuous agent operation independent of local machine uptime. Additionally, the workflow integrates services like Clerk for streamlined user authentication and billing (noting an “unlimited apps” deal), and DNSimple for programmatic DNS management, including an agent-friendly CLI for debugging.

The core methodology now emphasizes a conversational, high-level approach to prompting, moving away from prescriptive instructions to guiding the agent’s understanding. Key practices include using voice-to-text for initial “thought dumps,” critically analyzing the model’s textual outputs over just code, and initiating new threads for each distinct task to avoid context dilution. The workflow actively leverages visual aids like screenshots for problem definition and empowers agents with verification tools, notably Codeex’s computer use feature, to ensure task completion. This shift champions simplicity, minimal skill usage, and a focus on an “agent-native app substrate for tiny fullstack apps” philosophy. Furthermore, the developer advocates for an agile PR flow, utilizing agents for conflict resolution and evaluating branch relevance, while critiquing traditional SSH/terminal-based agentic coding as inefficient. The overarching theme is to “let go a little bit” and build in a way that maximizes the AI’s conversational strengths, ensuring the AI understands the developer’s intent and preferred building style.