LLM assisted development has nothing to do with trust at all. Think of it as a genetic algorithm. The model makes a rough approximation as a first pass, then it runs tests or specs, gets feedback, and iterates on it. And each iteration gets closer to what you specified. The better the specification that you provide is, the better the end result will be.

Meanwhile, context isn't some abstract thing. It's the amount of tokens the inference engine can hold in memory. With the memory constraints you have, it's going to be tiny, which means the model isn't able to hold the whole solution in the context even for a small thing like a puzzle generator which will be thousands of lines of code.

Good news is that pretty much any problem can be broken down into a series of steps. And then if you structure things well, each step can be treated as its own isolated problem which doesn't requite much context. With the example of a crossword puzzle, what you'd want to do would be to create an outline of the shape of the solution yourself, how the pieces fit together, and what the end result looks like, create tests for the functionality, maybe even files with stubs for the functions you want if you're working with a small model. Then you run an agentic harness like opencode where the model can iterate on the solution, run tests, and get feedback on what it's doing wrong. You create the scaffolding for the solution and you own the contract, while the model fills in the details for it.