Nuanced's founder says AI plan mode is becoming obsolete
The founder of Nuanced, a desktop coding app built earlier this year around the belief that planning would become the most important part of building software with AI, writes that the app's plan-mode approach failed, and that plan modes more broadly are losing their purpose. He says plan modes historically served two jobs: giving an AI agent instructions precise enough to execute, and helping the human understand what was being built. He argues the first job is rapidly becoming obsolete as models get better at exploring a codebase and making reasonable assumptions on their own, shrinking the need to spell everything out. The second job, he says, still matters, but plan modes are the wrong tool for it, especially as the number of parallel coding agents running at once grows from a handful to hundreds. He traces Nuanced's failure to several causes: he had conflated the act of planning with a fixed planning document; models improved to the point where much of that document became unnecessary; and the AI-generated specs the app produced were hard to read, describing the writing as having a pacing and overly structured quality that made his eyes glaze over. When the team tried fixing readability with a feature called Spec Tour, which walked users through key parts of a spec instead of the whole document, it only added another layer of complexity on screen rather than solving the underlying problem. He also concludes that Nuanced's linear workflow (chat, disambiguate, generate spec, review, revise, approve, implement, review code) split up a process that needed to feel continuous, forcing users to prematurely finish thinking before they could start building, with no way to move back up that sequence once implementation began. He points to how OpenAI's Codex currently works as evidence the boundary between planning and execution is dissolving into a single loop of understanding, acting, inspecting, clarifying and adjusting, rather than a distinct upfront planning stage. He calls turning the plan into a static artifact, instead of designing a process for better human understanding, the biggest mistake he made. He also says Nuanced's split between a plan mode and a build mode was awkward, requiring users to consciously decide whether a task deserved planning and then remember to switch modes, adding cognitive load that a plain chat interface avoided. He closes by saying the underlying problem of helping people stay oriented as increasingly capable agents change a system, especially as their numbers scale from five to hundreds, remains unsolved.
Key facts
- Nuanced was a desktop coding app built around the idea that AI planning would become central to software development.
- Its founder says plan mode's purpose of writing precise agent instructions is becoming obsolete as models improve, while the purpose of helping humans understand their own systems still matters but is served poorly by plan modes.
- AI-generated specs were hard to read due to their pacing and structure, and a Spec Tour feature meant to fix this only added complexity rather than solving the problem.
- Nuanced's linear workflow, from chat to spec to implementation, forced users to stop thinking before building and offered no way back, which the founder calls artificial and forced.
- He points to OpenAI's Codex as an example where the boundary between planning and execution is collapsing into a single loop, and says the deeper problem of human understanding as agent counts scale from five to hundreds remains unsolved.
Why it matters
This is a first-person account from someone who bet a product on AI planning becoming a first-class workflow step, and watched that bet fail as underlying models got good enough to close the gap themselves. It captures a broader shift in AI-assisted coding: as agents get better at exploring codebases and acting autonomously, the value of an explicit, human-authored planning stage shrinks, while the harder problem of keeping humans oriented as many agents work in parallel remains open.
Who it affects
Developers and teams building or using AI coding tools with explicit plan or spec modes, such as Claude Code CLI, Conductor and Codex, and anyone designing interfaces meant to keep humans in the loop as coding agents scale from a handful to hundreds running at once.
How to use it
The piece offers no product, price or release, only a founder's retrospective; the practical takeaway is a design lesson: rather than building a persistent planning document as a rigid artifact, interfaces should support an interleaved loop of understanding, acting, inspecting and adjusting, closer to how Codex currently operates.
How solid is it
This is a single founder's first-person account of his own product's design failure, backed by specific details about Nuanced's plan mode, build mode, and the Spec Tour feature it shipped; it is a personal essay and opinion piece rather than an independently verified case study, and no user numbers, dates, or details on Nuanced's current operating status are given.
Risks and caveats
The argument rests on one person's experience building one app and reflects his own conclusions about where AI coding interfaces are heading; it should be read as a considered opinion about a fast-moving space rather than an established industry consensus, and terms like 'better' models or 'good' at autonomous action are not tied to specific model versions or benchmarks in the piece.
“I think the biggest mistake I made was turning the plan into an artifact instead of designing a process for improved human understanding.”
— Nuanced's founder