Oh, so I am needed after all!

I’ve been playing around trying to get GLM-5.3 to build me a suite of C daemons for a personal project. I added some guidelines that should direct the orchestrator (Sonnet 5.5) to abstract common functionality into a statically linked library interface.

After requesting a few shared modules to consider for the archive I spotted that no archive (.a) file was being produced. Looking a bit deeper the Makefiles of the binaries we were building were actually individually compiling the source files and creating their own objects for linking. That’s like no library I’ve ever seen before. So I instructed GLM-5.3 to create a proper archive.

This is where things got interesting: GLM-5.3 claimed it would lead to object file bloat to link the same library to all the daemons if those objects weren’t being used. But this is false, this is literally what ld does all day long: find the objects you want, link them to the ones in this module and discard the rest.

I challenged the model on this and it told me this was a good spot. Yay me. The concern, of course, is whether it made the same mistake in other areas that I only half-understood (such as security). This is the first time I’ve detected something that is clearly untrue in GLM-5.3’s output.

Obviously I’m not working at the frontier here, that’s deliberate, so I’m trying to understand something about the shortfall, because if I can understand that then I can build equivalent software for less - because at the moment I think this is where the future is. For the record, I’m not sure a frontier model would’ve performed better, I would have hoped so, but it’s hard to run this sort of iterative experiment in parallel.

The conclusion? There’s a lot of talk about planning at the moment, with the removal of plan mode from Claude Code. I think I get it, in actual fact the planning does fall short of what’s needed, it’s insanely detailed in some ways (which is good - details matter) but those details are for me often part of the design limitations of the chosen solution path. I think I need a planning skill which explores the problem, so that comes next.