Comment by sinsudo
7 hours ago
Just an anecdote: I was designing a project for LLMs onboarding and orchestration. Claude chose to read only the first 40 lines of each file. Later, in another session, looking for causes of low quality result, Claude detected the fault and changed the code to perform an AST analysis, so now the analyzer takes documentation lines and functions signature (input/output) as input.
Claude's initial approach was really poor. One has to wonder how many times Claude code has to be modified/reviewed for improvement, or whether it is possible at all to make good code with it.
Edited: Generalization: Claude can fix a localized, identifiable poor decision (e.g., "only reading first 40 lines") because the fault is discrete and traceable to one piece of code.
But real software quality problems often arise from many small, individually reasonable decisions that collectively produce bad outcomes. No single one is obviously "the fault." In that scenario, a tool that generates low-quality building blocks piecemeal may never converge on good code, because each piece seems fine in isolation.
I think it's taught to look at source code through a peephole for the sake of context preservation, but I feel like this could be a good use-case for some sub-logic or even a full sub-agent. Like, here sub-agent, you skim that file and tell me a summary, and highlight any areas related to X and Y so that I can look at them in my main context. You can also periodically observe the main work stream and interrupt me if you realise that something in the file you're thinking about is relevant to what I'm working on or might change the direction of what I'm doing.
I think what you suggest is like a local second order approximation, that can help. But, I think that the real problem is a global one, is about architectural taste, how the many local pieces interact and their friction. Currently that demands human expertize.
> I think it's taught to look at source code through a peephole for the sake of context preservation
Yes, to a (real) fault. Less than one in fifty times it ignores an instruction or piece of data in a file has it seen the instruction or data before ignoring it. The other times, it's done this sampling nonsense.
Results are night and day using the 1M token models and reading the full files.