Plan mode is dead
The author of an AI coding app reflects on why its "plan mode" failed, arguing that while planning is crucial, rigid, AI-generated plans are ill-suited for modern, iterative AI-powered development workflows. Hacker News debates whether this signifies a fundamental shift in how developers interact with AI or if it's merely a symptom of evolving tooling and the enduring need for human oversight.
The Lowdown
The author, jmvldz, recounts the story of their desktop coding app, Nuanced, which was built around the concept of "plan mode" for AI-assisted software development. Despite a strong initial belief that planning would become the most critical component of building software with AI, Nuanced ultimately failed, leading the author to conclude that plan modes, in their traditional form, are no longer useful.
- Nuanced aimed to help humans maintain a coherent mental model of software systems rapidly changed by AI agents, as existing interfaces hadn't kept pace with accelerated code generation.
- Historically, plan modes served two functions: providing precise instructions for agents and aiding human understanding. The author argues the first is becoming obsolete due to increasingly capable models.
- The failure of Nuanced stemmed from several factors: conflating "planning" (the process) with "a plan" (a rigid artifact), the rapid improvement of AI models, the difficulty of reading verbose AI-generated text, and disrupting the continuous nature of planning and building.
- The author observed that the development process is inherently iterative, moving through understanding, acting, inspecting, clarifying, adjusting, and acting again, rather than a linear waterfall of planning then execution.
- The explicit separation of "plan mode" and "build mode" in Nuanced added cognitive load, forcing users to prematurely "finish thinking" before starting to build.
- The core challenge of human understanding in systems modified at high velocity by numerous agents remains, and a lengthy AI-generated document isn't the right interface for it.
The article concludes that while the need for human understanding and navigating complexity is constant, the interface for planning must evolve from static artifacts to dynamic, iterative processes that integrate seamlessly with AI-driven development.
The Gossip
Planning's Perennial Paradox
Many commenters, including the author, agree that while the *concept* of planning is indispensable for complex tasks, the traditional 'plan mode' as a distinct, rigid phase or AI-generated artifact is indeed dead or dying. They suggest that modern LLMs are intelligent enough to implicitly understand the need for planning through conversational prompts or that the planning process should be continuous and iterative, not a separate, upfront document. Some note that 'plan mode' was often just a specific prompt and that better models obviate the need for explicit enforcement.
Human Comprehension and Technical Debt Conundrum
A significant thread expresses deep concern about the potential for AI-driven development to erode human understanding of codebases and accrue massive technical debt. Commenters worry that developers are losing the ability to grasp the 'why' behind AI-generated code, leading to bloated, unreadable systems. While some propose using AI to *explain* code, others emphasize that true understanding comes from critical engagement, not passive consumption, and that neglecting this will have severe future consequences.
The Art of Agent Orchestration
This theme explores practical, evolving strategies for integrating AI into development workflows. Many users describe custom setups, where they explicitly prompt agents to plan in markdown files, use multi-agent systems (supervisors, minions, validators), or iteratively refine plans before execution. The discussion highlights the importance of guardrails, cost management (using different models for planning vs. execution), and the challenge of moving from simple one-shot tasks to complex, stateful projects.
Tooling Trials and Tribulations
Commenters share frustrations with current AI coding tools, including models misunderstanding intent despite advances, generating verbose and painful-to-read text, and the high cost of tokens. There's debate about whether proprietary tools like Claude Code are adequately addressing user needs or if open harnesses and custom solutions offer more flexibility and control. Some lament the removal or deemphasis of features like 'clear context on plan accept,' which they found valuable for managing context and costs.