ChatGPT is good to discuss technical aspects, data structures.
Claude Sonnet and Gemini 2.5 Pro are good with coding.
If the indicator is complex, get the AI to code simplified indicator first than add features after fully discussing all the aspects before hand. Sometimes it helps to start with fresh context.
If you are getting compile errors - feed them to AI. If AI makes incompatible code - provide it with docs.
For runtime errors - you can describe the problem, ask it to add debug to console (which you can feed it to AI) or have AI analyze the screenshot.