I run a stack with the Qwen API inside VSCode via the KiloCode plugin and the regular GPT 5.1 Thinking inside ChatGPT. Based on previous experience with both Qwen and Gemini 2.5 Pro, using GPT 5 and 5.1 as the planner or the architect outside of the development environment significantly outperforms directly prompting the agent that will write the code.
In my most recent project where I’m porting over a project from Figma Make to Encore.ts, I’ve found that what I’m really engaged in now is not only assisted development, but a game of meta-coding.
For each task, I have to determine the right approach and which model to give the task to.
For a small task, I generally feel comfortable giving it to Qwen directly to execute on.
For larger architectural tasks that may need significant refactoring across the codebase, I need to consider whether I should give it to GPT 5.1 instead.
It becomes a matter of which model do I trust to get this right and to not run in circles writing and re-writing the same code over and over again.
I also need to consider how many steps in the analyze-specify-plan-implement-test process I need to run for each task.
And before that, how much context do I need to generate to ensure that either model knows what it is doing.
Then, I need to consider whether I can go ahead and give the direct requirements to the model - do I actually know what needs to be done, or do I need the model to determine that? And in that case, I should rather define a high-level goal, a current problem, or an expected outcome, and let the model define the practical steps to reach the end state.
So meta-coding is a matter of deciding which model do I trust with this task, and do I approach the task directly or indirectly.