Developers explore new ways to use LLMs for coding

A recent discussion on the Hacker News front page highlighted a growing frustration among developers who rely on large language models (LLMs) for coding. The thread, posted by a user who uses Claude Code and Codex, describes the experience as “a bicycle that just brakes abruptly every couple minutes.” The commenter notes that the current prompt‑response workflow interrupts the flow state that many programmers achieve when writing code by hand, forcing repeated cycles of stopping, reviewing, and re‑prompting. They also suggest that a “tab model”—where the model can maintain context across multiple, parallel code snippets—might better support continuous development.

Several startups and research groups are exploring alternatives to the traditional prompt‑response paradigm. For instance, Tabby, a GitHub‑based extension, allows developers to keep multiple code contexts open simultaneously, reducing the need to re‑enter prompts. Other projects, such as CodePilot and KodeAI, are experimenting with conversational interfaces that treat code editing as a multi‑turn dialogue, aiming to preserve context over longer sessions. Academic initiatives like OpenAI’s Codex‑Chat and Microsoft’s GitHub Copilot Chat also attempt to embed a more persistent context layer, though early user feedback indicates that the transition to a truly seamless “tab” experience remains incomplete.

While the community remains divided over the best way forward, the conversation underscores a clear demand for more fluid interaction models. As LLM‑powered coding tools evolve, developers and researchers alike are keen to move beyond the stop‑and‑prompt cycle toward interfaces that support sustained, uninterrupted work. The ongoing experimentation with tab‑like contexts and conversational coding suggests that the next generation of AI assistants may finally bridge the gap between rapid idea generation and deep, uninterrupted coding flow.