A developer who built a desktop coding app around planning with AI has declared that plan modes are dead. The app, called Nuanced, was meant to help people keep track of what they were building while AI wrote code. It failed, and the author believes that reason reveals something about where development is going.
The problem Nuanced was built to solve
This year’s opening months saw the author argue that planning would be the defining skill of building software with AI. The case rested on a single observation: AI can produce thousands of lines of code in minutes, while the tools people use to work with it have not come anywhere near matching that speed.
Maintenance was the issue. Before you’d even thought through what you were building or why, you inherited a huge pile of code. That made it hard to reason about behavior and debug incorrect assumptions that had become code.
A dopamine reward came from the simple act of coding, yet it concealed the true labor of working out why building something matters, whether it matters at all, and judging product, design, and infrastructure decisions. The author frequently produced a product without ever deciding anything about it first. Architecture left under-specified meant the agent stepped in to fill the gaps, creating problems later on. Errors about intended behavior and design slipped through several files far beneath the surface of chat, and could be easily missed.
The experience left the author feeling mentally disconnected, a condition that made it harder to verify the accuracy of anything produced.
What plan modes actually do
Plan modes historically served two purposes:
- They specified instructions that were sufficiently precise enough for an agent.
- They helped humans understand what they were building.
The writer believes the initial reason for using plan modes is quickly becoming unnecessary thanks to improvements in models. The second reason, meanwhile, matters more than ever, though plan modes remain the wrong abstraction for it, especially as the number of parallel agents grows.
Why existing plan modes didn’t work
Each of the tools the author tried — the Claude Code CLI, Conductor, and Codex — lacked the right abstraction for planning. As conversations progressed, plans were ephemeral blobs of text that vanished into the backscroll, and none of the trio had a solution for that problem.
The cut-and-paste workflow felt awkward and made it difficult to work through an idea while keeping track of the current plan. What the author wanted was a home for plans, a place to ground them in the overall workflow, and to turn plan mode into a persistent, living document.
What Nuanced actually did
Nuanced let users spin up threads, where each thread was a chat conversation. You’d talk through what you wanted to build, and the system would surface ambiguities and decisions that required your input. Together you’d arrive at a persistent plan before implementation began.
Nuanced would put that plan into effect, ensuring the generated code stayed true to your requirements. The concept called for a pipeline that moved from intent all the way through to implementation, review, and verification, in one continuous process.
Rather than a standard coding app, the author viewed it as a prosthesis for the human mind — or an aid for their ADHD mind — with an emphasis on instruction as much as on keeping track of what was going on.
Why it failed
The author’s ideas about gaps in existing tools and how the software development lifecycle was changing weren’t incorrect. The implementation simply didn’t deliver the solution expected.
The main reasons were:
- Conflating planning with a plan
- Models got really good
- No one wants to read AI-generated text
- Separating planning from building in a way that was disruptive
Space to think before a project starts is not the same as actually writing down a plan. The idea was that keeping a thorough record of that early thinking in a large, organized document would be valuable as the project changed over time. Instead, early users wanted to see the results sooner rather than spend time reading through long records of planning.
The new shape of planning
The point being made here is that making plan modes serve as a tool for people to grasp what they are building matters more than ever. Yet the way these modes are constructed does not hold up, particularly as the count of parallel agents grows.
The existing plan modes lacked collaboration. The author wanted to anchor their plans in the larger workflow, give them a permanent home, and transform them from fleeting scraps of text into lasting documents.
A journey from intent through implementation, review, and verification would define the ideal workflow. It would serve as an extension of the human mind, designed as much to guide as to keep track of progress.
What this means for developers
The author’s conclusion is blunt: plan modes are dead. The problem was not the idea but the execution. The implementation didn’t deliver the solution expected.
No error occurred on account of faulty reasoning. The argument over the shortcomings of current tools and the shifts taking place within the software development lifecycle was accurate.
The abstraction that was missing was the one that would have helped people understand what they’re building. Plan modes, which are typically used to organize work, are not suited to tracking what happens when the number of parallel agents grows.
| Plan modes historically | Nuanced tried to fix |
|---|---|
| Specified instructions precise enough for an agent | Persistent plans before implementation |
| Helped humans understand what they were building | Collaborative home for plans |
| Ephemeral blobs of text in chat backscroll | Persistent documents |
| Not collaborative enough | Grounded in overall workflow |
The author felt mentally detached, in a trance-like state, as coding apps such as Conductor and Codex enabled running more agents simultaneously. It became more difficult to confirm whether the generated results were accurate.
Source material: “Plan mode is dead,” aymannadeem.com.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.