Do We Need Plan Mode?

The article "Plan mode is dead"
https://www.aymannadeem.com/artificial/intelligence,/developer/tools/2026/09/24/plan-mode-is-dead.html

Ayman Nadeem, founder and CEO of Nuanced, published a postmortem of her own product, which was built around the idea of an LLM plan as a persistent spec document. That approach was precisely what failed. Long AI-generated text is difficult to read. Her follow-up post, "AI coding interfaces make approval easy and understanding hard", continues the same argument: the problem isn't planning itself, but the fact that the interface makes approval easier than understanding.

The author explicitly acknowledges that planning is important. The takeaway here is that a particular abstraction doesn't work: a separate mode built around a lengthy AI-generated spec. The linear plan → approve → implement workflow doesn't fit well with an agile approach or exploratory work, where every step changes your understanding of the task. The thinking process is valuable; a huge generated spec is not.

The Main Counterargument from Claude Code
https://news.ycombinator.com/item?id=49850929

On September 23, a day before the article was published, Thariq from the Claude Code team wrote on X that the team was considering removing plan mode and repurposing Shift+Tab for effort levels. This wasn't a final decision: the official documentation still includes /plan and plan mode.

In the thread, bcherny (Boris Cherny from the Claude Code team) replied: "plan mode has always been a prompt" — it's just a reminder included in every message, not a separate set of tools. They haven't changed it because doing so would break the prompt cache and make sessions more expensive. It's important to note that, from a product perspective, it's not just a prompt: it's also a permission mode with an editing gate. In sessions where bypassPermissions is available, that gate doesn't apply, so plan mode shouldn't be considered a sandbox.

He says he personally stopped using the mode with early versions of Fable and Opus 5.5, because the model started planning interactively and iteratively. But he also adds: "plan mode isn't going away"; at most, Shift+Tab might be reassigned for people who don't use the mode. Still, everything depends on the model, budget, and codebase. With cheaper models and unfamiliar repositories, plan mode can still save time.

The practical consensus in the HN comments isn't "don't plan", but rather "don't pay for a lengthy AI-generated plan that you never read". For risky or multi-file changes: write the plan to a file, use a stronger model for planning, a cheaper one for implementation, and have a separate review gate.

I often use plan mode specifically as a read-only mode (some harnesses have an Ask mode for this purpose) when I need the model to leave everything untouched and simply gather information. I also use the plan (Opus/Sonnet) → implementation (Sonnet/Luna/Chinese models) approach, but I don't switch modes to do so.

#claudecode #codex