Plan Mode is the most slept-on feature in Claude Code – and we just made it a lot better. It dramatically improves your one-shot success rate. Fewer iterations, less hand-holding, better final results. Shift + Tab twice to try it.

Plan Mode now spins up multiple subagents in parallel to explore your codebase and generate plans from different perspectives. Then it asks clarifying questions before committing to an approach. The result: more comprehensive, better thought-out plans before Claude writes a single line of code.

For example: when I was building the multi-subagent UX you see in the video, Claude generated plans from two perspectives – one hacky and fast, one architecturally sound. The main agent balanced the two based on my intent and gave me something that worked really well. This is Plan Mode at its best.

Claude now writes plan files to your filesystem, and you can edit them by running /plan open. Any changes you make are automatically synced back to Claude's context, so you can tailor Claude's plan to your exact requirements if it's not perfect

Tip: You can set Plan Mode as your default with /config More details:

Unfortunately it’s not better because the plan now will be created based on what haiku (Explore) finds and Opus only write the plan based on that and haiku is not so smart so most of the findings will be incorrect

Explore agents are only used to explore the code and understand how it’s structured. The main agent will then read some of the code itself, and spin up “Plan” agents (not haiku) to do more comprehensive planning depending on the complexity of your task

hmm something is wrong then, opus doesn’t seem to spin up the plan agent at all

Fell over these explore agents, felt really like magic! Plan mode was the thing I missed the most when using Antigravity lately, since that thing always eagerly jumped to writing code.

claude --dangerously-skip-permissions needs to also apply to plan mode. it's annoying that even with bypass permissions mode on, plan mode constantly asks if it can do every little thing.

I've been implementing my own custom plan mode since the beginning of Claude Code and avoided "Plan Mode" because it hasn't respected the --dangerously-skip-permissions flag. I didn't want to sacrifice autonomy for better planning. Is it possible to enable it without requiring us to approve actions? I guess I could have an alias that starts claude with a flag that passes in all possible approvals. Is that the best way for now?

plan mode is my default mode when I start working on something

Ok real talk I built my own flow but it's based on the CC flow of Plan mode -> Explore/Plan subagents -> Master makes the plan. It's really really simple, yet works out so well.

will you be bringing back the opus plan /sonnet build model choice?

I always switch to plan mode when I am afraid that Claude Sonnet/Opus gives me a few options, and then ... without waiting for an answer from ... just moves on implementing the option it likes best. 😬 Even happens with DROID as CLI, somit must be in the model already !

it is fantastic. but using plan mode before work can't be considered one-shotting IMO :)

fair!

Strong agree. Gets the most bang for the buck. Doesn't need complicated subagent or skills setup, just use Plan mode and you're good to go for ~80% of the use cases.

Explore agents are great. New planning system is annoying though when it finishes plan and I ask for something additional, then the agent wants to update the entire plan. Updating the plan reprints the whole thing again when I just want to know about new task.

when you re-enter plan mode are you trying to iterate on the original task, or start a new unrelated task?

It's an additional feature after the main plan - /clear would waste tokens having Claude get up to speed on the original work. The new task is relatively small compared to the plan, but Claude wants to revise and reprint the whole plan. I just want an update on the new task.

I probably did use plan mode on new feature, but it's usually because I want to discuss details before implementing the new thing. Maybe there could be a /clearplan slash command that keeps context but marks existing plan as complete? Just an idea.

`CLAUDE_CODE_PLAN_V2_AGENT_COUNT` 👀

I think u should still add an option, something like "plan with opus / execute plan with sonnet". This option is not available at the moment, is it? I guess you would save some tokens this way and also speed is probably higher with sonnet? Just wondering!;) But claude code is TOP

Plan mode is great, I use it a lot. But it always requests for additional permissions: find, cat, ls, grep, use context7, psql, etc.. - each time something new. In the new repo this becomes a nightmare. Why just don't let it use all read commands? Or am I missing something?

I want to tell you the thing I hate about plan mode. It’s static. Sometimes it will ask me a question and I can tell it doesn’t get the issue. Then I have to go through all the questions. I want the questions to update as I answer

Best feature on the app. Makes Codex less enjoyable to use.

can confirm! nitpick when during the exploration bot is running a permission dialog comes up, the TUI reacts extremely sluggish to let me move to the right selection / provide different text

Do you guys ever get tired of cooking?

Claude code improves so much better the last 4 - 6 Weeks. It' s awesome. Very good work here!

Yea I always recommend starting any new feature or idea in plan mode in my Claude Code guide. I’m glad it now saves the plan to a markdown automatically

actually i have the plan mode as a default for new chat🤝

This is really magical, I just love something about agents and humans connecting and agents not ignoring and building something of it's own, rather cross checking it's own thoughts.

Maybe this is why I’m not getting good results. Will def try in my next idea.

Great feature! I'm using it for a time now and it indeed increased my work flow with better output and 'one shot' building. Also work even better with OpenSpec

Yep. I always do plan before implement.

thank god plan mode exists, thank you guys

I max out my $200 subscription every week and I'd say about 75% of my spend is in plan mode. It's rocket fuel

well done, just gave it a go and it managed a complex refactor like a champ.

Stupid Belgium is here, but not Belarus. Are you idiots there?

Is it cursor integrated? For example, when i choose Plan Mode in Cursor, is it going to run this plan mode of the model (like a tool) or Cursor’s Plan Mode?

Is there any way Claude is able to know the path of an image file I drag into the terminal? It’s bizarre that claude can “see” the image but not know its location. This simple change would be a huge quality improvement as many of the skills and workflows I want to use involve referencing an image dragged into the terminal but Claude currently has no idea which file it is!

Can it be invoked in the ios app, Code mode?

Love plan mode, can I switch to opus plan sonnet build?

what’s the use case for this? you get as much opus now as you did sonnet previously on subscription plans

Hey your support confirmed that the session limits weren’t increased just the weekly, so me running into session limits using just opus makes sense. And I think there would definitely be a use for opus plan sonnet build.

Plan Mode’s been good for bigger refactors. The step-by-step plan cuts so many back-and-forths.