So far the follow-up questions have all been really good design/engineering questions that I appreciate.
I saw a Youtube video that showed Sonnet-4.5 is much better than GPT-5 at making plans in Cursor and tried it out. I 100% agree! Sonnet plans tend to include what files to touch and details about the implementation. grok-code-fast-1 could nail the implementation on my last one
Iâm already using it in another repo. without workspace.
I loved the PLAN mode too. Previously I had to tell the AI to just plan and not to implement anything. But sometimes, when I was revising the plan, it started to implement some changes and I had to revert. Now, using the PLAN mode, I have a piece of mind that it wonât happen. Also it seems that in the PLAN mode, it really try to understand the codebase thoroughly so I have an easier time writing my planning prompt. Good job Cursor Team!
Make a plan, clarify specifics, ask to be explicit, build.
Realize build is missing something, return to plan, adjust, build.
This is a nice approach because you can keep the agent consistently corralled to the plan, also if you ask the smarter planning agent to include better snippets that you want the implementation agent to use so theyâre not just making it up as they go.
I really enjoy using the plan mode as well.
It will be nice if the build can be split up for each step instead so I can walk it through.
Sometimes my plan is bigger than usual, and I will like to review the codes for each step instead of a full build.
I use in proper order
Ask - I being with carrying out discussions . Share my thoughts , corrections
Plan - Once I an convinced, I move to plan mode. The plans are comprehensive this way. Continue discussions / corrections
Agent - In the end move to agent mode.
Have been a good experience so far .
what models?
It seems optimized on Sonnet 4.5 (read somewhere) - no sure if thinking.
I made relatively simple plan. Sonnet offered either EPPlus or ClosedXML libraries to be used. I told I want ClosedXML, but it is not able to update the plan. It says something about file isnât returned by the tool.
Two issues I have been encountering:
When creating plans they are empty every time. I see the chat event âCreated Planâ printed, and it is attached to the chat thread, but no preview graphic post.
I click on the attached plan and it only has to-dos, but content is empty.
I have to tell the agent it is empty and to regenerate it. THEN I see the plan graphic with the skeleton-style-glow-animation on the title while working and a âread detailed planâ link.
UX confusion: When I click âSave to workspaceâ it saves a copy with a short-UUID-suffix under /.cursor/plans/..., but this is not the plan the agent is using in the chat. It would make more sense if plans were either always saved, or it replaced the one in chat with this one. Right now I end up with both of them open and they both update.
Bad file @: I donât recall at what point this happened, but the plan gets dropped from the chat tab and @ still finds it but when accepting the autocomplete nothing is displayed. I think Cursor does this for any deleted file until it reindexes (another bug, but minor). I can still @mention the saved plan with the id suffix and continue on working with the plan that way.
I do the same.
However, this step I do with chatgpt.. then ask it to give me a prompt to use in cursor to plan what I want.
Then I plan with cursor.
This is working really well atm, itâs kind of crazy.
Plan Mode is an awesome concept, but this implementation is barely working for me.
The Save to Workspace button doesnât save anything to workspace, and Cursor doesnât seem to reliably use the plan to track its progress but instead Claude goes and creates several new markdown files to describe its progress.
I really want a single canonical plan that gets updated.
Why was 4.5-haiku removed from the plan mode models? It was the best combo for me 
The empty plan and models having trouble updating the to do lists is really slowing this down and burning tokens for no reason. I would love to know if this has been reproduced by the dev team.
I donât bother with Plan mode. Instead I say âWrite a plan in this markdown fileâ. Then I can edit the plan, and tell the model to execute the plan or parts of it.
Some items Iâve noticed
Can we make it more like Google Antigravity where I can comment on line/text, ask questions, give feedback?
I canât easily reproduce this, but Iâve had 2 instances of using Plan Mode where once I hit build, the plan got truncated (only the first couple of sections existed - though I do not know if this was the agent rewriting the plan as a summary when starting to work or if it was literally just dropping the content after a certain line), and since the one saved to the project is like a hard link, that one changed as well. This caused the agent to make a lot of implementation mistakes because it was only going off its context from making the plan I assume, and so only had detail for the early and last few prompts.
Iâll try and get in to the habit of copying the markdown before pressing build I suppose.
This is a very costly issue as I spent over an hour working on my last plan.
