Earlier quoted context omitted.
Copilot doesn't use the full context length. Write scripts to dump relevant code into claude with it's 200K or the new Gemini with even more. It does much better with as much relevant stuff as you can get into context.
But I don't want to write additional scripts or do whatever additional work to make the 'wonder tool' work. I don't mind an occassional rewording of the prompt. But it is supposed to work more or less out of the box, at least this is how all of the LLMs are being advertised, all the time (even the lead article for this discussion).
(tree Source/; echo; for file in $(find Source/ -type f ) ; do echo ======== $file: ; cat $file; done ) > /mnt/c/Users/you/Desktop/claude_out.txt #claudesource
You can also do stuff like pass in just the headers and a few relevant files
(tree Source/; echo; for file in $(find Source/ -type f -name '\*.h' ; echo Source/path/to/{file1,file2,file3}.cpp ) ; do echo ======== $file: ; cat $file; done ) > /mnt/c/Users/you/Desktop/claude_out.txt #claudeheaderselective
Put the command itself into claude too and tell claude itself to write a similar one for all the implementation files it finds it needs while looking those relevant files and headers.
If you want a wonder tool that will navigate and handle the context window and get in the right files into context for huge projects, try claude code or other agents, but they are still undergoing rapid improvements. Cursor has started adding some in too but as subscription calling into an expensive API they cost cut a lot on trying to minimize context.
They also let you now just point it at a github project and pull in what it needs, or tools build around the api model context protocol etc. to let it browse and pull it in.