Evolution of an AI-Assisted Development Workflow

The Shift from CLI to Desktop Apps

The author has moved away from terminal-based agent workflows (like Claude Code CLI) in favor of dedicated desktop and web-based IDEs. While CLI tools were once the standard for agentic coding, the author argues they are fundamentally limited by poor image handling, session management issues, and the friction of managing terminal multiplexers (tmux/zellij). Desktop applications like T3 Code and the OpenAI Codeex app provide a more cohesive experience, offering better support for image pasting, thread management, and persistent remote sessions.

Remote Development and Infrastructure

A core component of the author's current workflow is decoupling the development environment from the local machine. By hosting T3 Code on a remote server (or a Mac Mini) and accessing it via a browser or desktop client, the author maintains state even when offline or switching devices. This approach solves the "laptop-half-open" problem, allowing agents to continue working on long-running tasks without interruption. The author highlights the use of Tailscale for secure, low-friction network access to these remote instances.

Context Management as a Strategy

Rather than relying on complex "plan mode" prompts, the author now focuses on aggressive context management. A key technique involves giving the agent access to relevant existing codebases by cloning them into a scratch directory. This allows the model to reference working implementations (e.g., authentication services) directly rather than relying on potentially hallucinated documentation or abstract descriptions. This method significantly increases the reliability of generated code.

Model and Tooling Preferences

The author currently favors GPT-5.5 due to its generous usage limits and performance, having largely moved away from Claude models for primary development tasks. While the author acknowledges the utility of the Codeex app, they advocate for T3 Code—an open-source project they co-developed—for its stability, remote-hosting capabilities, and lack of "bullshit" credit-based usage limits that have recently plagued other proprietary agentic tools.

Notable Quotes

• "I really, really don't like doing agent coding in terminals anymore... A good desktop app for coding will shit all over a CLI any day."
• "I use images a lot in my prompts. The fact that they kind of almost work in traditional CLIs and then don't work at all over SSH is insulting."
• "Context management really is the name of the game for getting these things right."
• "I still can't believe the pasting images over SSH thing is as absurd as it is... I will not support that mental illness."

Actionable Insights

• Stop using terminals for agentic work: If you are fighting with SSH, tmux, and image uploads, switch to a dedicated agentic IDE.
• Use remote hosting: Host your development environment on a machine that stays on (Mac Mini, VPS, or Replit) to ensure agent tasks continue running when you close your laptop.
• Clone for context: When building new features, clone your own relevant past projects into a local scratch directory to provide the agent with a concrete, working reference implementation.
• Secure your remote access: Use Tailscale to connect to your remote development machines without exposing them to the public internet.