I’m using the GPT-5 model in Cursor IDE, and overall I really like its performance.
However, I’ve noticed one issue: I can’t get it to create its own task list and then execute it step-by-step.
This exact feature works perfectly with other models (e.g., Sonnet, Opus, Gemini), but for some reason GPT-5 simply refuses to work in this way.
It would be great if this capability could be added or fixed, as it would significantly improve how complex tasks are organized and executed.
That is true, GPT does not create task list and it is one of the main reason why it fails at executing on task that require writing more than 7 lines of code or building several features. I hope the cursor team sees this and fixes it cause it is the one annoying thing I find about it. Also it doesn’t listen sometimes and you have to strongly ask it to read the entire file or at least more than 6 lines of code before coming up with a solution.
im in the same situation here. Instruction following is broken. It tends to forgot rules, proposing weird replatforms all outside stablished technical and design guidelines… it was working like a charm before GPT5, still works great with any other model.
And fun fact, this not being able to follow instructions is happening since Friday. On thursday it was wonderful, honestly almost 0 complains, since friday GPT 5 in cursor became even worst than claude 4 at following instructions, is like completely blind.
Hi, yes the todos are not yet smooth with GPT-5 and we are working on it. Thank you all for the feedback!
Could you post a full separate Bug Report with more info Create Bug Report. More info helps us determine causes and fix issues.
Also add a Request ID with privacy disabled so we can look into the details. https://docs.cursor.com/troubleshooting/request-reporting
I’m not entirely sure what specific example you’d like me to provide here. In my experience, GPT-5 has neversuccessfully created a todo list when asked — even when I explicitly request it, it states that it will do it, but then doesn’t.
I can certainly create and share an example for you, but honestly, it seems a bit pointless since it fails to do this in every case I’ve tried. Still, I’ll prepare one and share it soon.
@LohiSoft thank you and I agree, the todo feature limitation is known and team is working on it.
Any other not well known issues should be filed as Bug reports so we can analyze them. The Request ID with privacy disabled helps us to see what went wrong within the request.
cant use todo in gpt5
cant use todo in gpt5
Linux
版本: 1.4.3
提交: 1.99.3
日期: e50823e9ded15fddfd743c7122b4724130c25df0
Electron: 2025-08-08T17:34:53.060Z
ElectronBuildId: 34.5.1
Chromium: undefined
Node.js: 132.0.6834.210
V8: 20.19.0
OS: 13.2.152.41-electron.0
hi @Jiawei_Gu thank you for the report. This a known limitation and the team is already looking into it.
Hi everyone! I’ve noticed that GPT-5 often gets stuck on simple tasks after creating a long implementation plan, either in text form or as a .md checklist.
It tends to stall after completing 1–2 steps of the plan, repeatedly summarizing what it has already done and what it will do next—without actually doing it.
Sometimes it writes 1–3 lines of code related to an old task, which are often unnecessary.
Usually, I ask it to “continue,” but that doesn’t work. When I give specific instructions to proceed, it rarely helps. It often takes 3–5 prompts before it moves on to the next step. This is hard and frustrating to work with, especially considering the request cost. It may spend 400k tokens just to produce summaries and outline next steps, or 1,000,000–1,500,000 tokens when calling tools to add only 1–3 lines of code.
I understand the issue is with the long context, which quickly fills its memory and causes it to slow down.
This behavior happens with GPT-5, GPT-low, and GPT-high. However, I’ve noticed GPT-5-mini doesn’t have this problem. My GPT-5 got “stuck” on a task, and I switched to GPT-5-mini in the same chat. It handled the implementation plan well and responded correctly to “continue” requests, even in a chat where the context was already 91% full.
I can see that GPT-5 requires additional tuning.