Plan mode is dead

The concept of plan mode, once believed to be essential in software development with AI, has become outdated. Nuanced, an AI-powered coding app, sought to address this trend by creating a desktop application centered around planning. However, the product ultimately revealed that plan modes are no longer as useful as they once were. Historically, plan modes served two primary purposes: providing precise instructions for AI agents and helping humans understand their building process.

While plan modes have become obsolete due to advancements in AI models, the need for humans to maintain a coherent mental model of a software system as machines change it faster than humans can inspect the changes remains relevant. The ease of generating code using AI has led to a greater dopamine reward but has also obscured the difficult work of understanding why building something matters, whether it matters at all, and rigorously evaluating product, design, and infrastructure decisions.

This has resulted in misunderstandings about intended behavior and design that propagate through multiple files, making it challenging for developers to reason about behavior and debug incorrect assumptions.

The experience of using coding apps like Conductor and Codex, which enabled the running of multiple agents in parallel, left the author feeling mentally disconnected and unable to focus, leading to difficulties in verifying the correctness of generated results. This disconnect, coupled with the lack of a clear, interpretable trace connecting user prompts, agent decisions, code, and product behavior, has led the author to seek a new abstraction for planning well.

The author sought a solution by encoding their dream workflow into a product that would act as a prosthesis for the human mind, particularly for those with ADHD. The product aims to guide development from intent to implementation, review, and verification, serving as a first-class primitive for development. The author learned that planning and a plan are not the same thing, and that preserving the thinking in a large structured artifact became less valuable as models got better at understanding large codebases and making reasonable assumptions.

The focus shifted towards providing a persistent document for planning, allowing developers to ground their ideas in a living, breathing workflow.