The context windows are only so large. Once you give it too much to juggle, it starts doing crazy shit.
Boilerplates are fine, they can even usually stub out endpoints.
Also the cheap model access is often a lot less useful than the enterprise stuff. I have access to three different services through work and even inside GPT land there are vast differences in capability.
Claude Code has this REALLY useful implementation of agents. You can create agents with their own system prompts. Then the main context window becomes an orchestrator; you tell it what you’re looking for and tell it to use the agents to do the work. The main window becomes a project manager with a mostly empty context window, it farms out the requests to the agents which each have their own context window. Each new task is individual, The orchestrator makes sure the agents get the job done, none of the workloads get so large that stuff goes insane.
It’s still not like you can say, go make me this game then argue with it for a couple of hours and end up with good things. But if you keep the windows small, it can crap-out a decent function/module if you clarify you want to focus on security, best practice, and code reusability. They’re also not bad at writing unit tests.
The context windows are only so large. Once you give it too much to juggle, it starts doing crazy shit.
Boilerplates are fine, they can even usually stub out endpoints.
Also the cheap model access is often a lot less useful than the enterprise stuff. I have access to three different services through work and even inside GPT land there are vast differences in capability.
Claude Code has this REALLY useful implementation of agents. You can create agents with their own system prompts. Then the main context window becomes an orchestrator; you tell it what you’re looking for and tell it to use the agents to do the work. The main window becomes a project manager with a mostly empty context window, it farms out the requests to the agents which each have their own context window. Each new task is individual, The orchestrator makes sure the agents get the job done, none of the workloads get so large that stuff goes insane.
It’s still not like you can say, go make me this game then argue with it for a couple of hours and end up with good things. But if you keep the windows small, it can crap-out a decent function/module if you clarify you want to focus on security, best practice, and code reusability. They’re also not bad at writing unit tests.
Something like speckit is necessary to make big, sweeping changes that continue past the context window
Interesting project, thanks for sharing