Planning
Using a model like o3 in ChatGPT can be useful to flesh out an idea and iterate before building anything.
Once you are happy with the plan you can even ask o3 for an initial prompt to put into your agent to start on the build.
Model Selection (for coding)
While there is an array of models to choose from. My two recommendations (as of today) are:
Gemini 2.5 Pro
This is the model with the fastest output and seems to be the better model for highly interactive graphical experiences (e.g games, 3D, simulations).
Even if your idea is more on the regular web UI side then it still might be best to start with this model due to the speed of output.
It also excels at high level thinking of about a concept and planning.
The downside of 2.5 pro is that it often likes to refactor and modify things unrelated to your prompt. Which might not be much of an issue with prototype code.
Claude 4 Sonnet
This model seems to have the most accuracy in adhering to a prompt and understanding what you are looking to change.
It also seems to be better at web UI design, so if you are looking to build an app that is more of a traditional web app this model might be a good choice.
The downside to Claude 4 Sonnet is that it has a much slower output when compared to 2.5 pro.
Larger models: o3, Claude 4 Opus etc
While the larger models might provide the best output they are much slower to run and potentially might use all of your Cursor request budget in a few prompts.
Workflow
Once you are in the thick of the build keeping the model on track and staging your changes as you are happy with changes should be your priority.
Stage changes (or commit) when you are happy with the output from a model.
If the model makes a small mistake, follow up with another prompt to correct it.
If it still is going way off track. Make use of the “Restore checkpoint” feature button found on your previous prompt to erase the models memory of the uncorrect path it went down. Then follow up with a prompt to include more context on what you want to achieve.