3 ms·

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 bet

by cma 2y ago

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.

hansmayer 2y ago
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).
- cma 2y ago
  LLMs are also primarily promoted through the web chat interface, not always magic wonder tools. With any project that will fit in claude/gemini's large context you use those interfaces and dump everything in with something like this: (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 Then drag that into the chat. 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 You can then just hit ctrl+r and type claude to refind it in shell history. Maybe that's too close to "writing scripts" for you but if you are searching a large codebase effectively without AI you are constantly writing stuff like that and now it reads it for you. 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.
  - hansmayer 2y ago
    No thank you for obviously good intent on your side, but I am not looking for scripting help here, nor am I a business type who does not code themselves. I just don't want do this when I am already paying for the tooling which should be able to do it themselves, as they already wrap Claude, ChatGPT and whatever other LLMs. And unless you're professionally developing with Microsoft stack, I'd advise to ditch the Windows+MinGW for Linux, or at the very least, a MacBook ;)
    - cma 2y ago
      The tooling you are paying for doesn't work with the full abilities of the context so you need to do something else. Doesn't matter what it's supposed to do or that other people say it does everything for them well, it works a lot better with as much in context as possible on my experience. They do have other tools like RAG though in cursor, and it's much quicker iteration, ultimately a mix of what works best is what you should use, but not just block stuff out out of disappointment with one type of tool.
      - hansmayer 2y ago
        I am lucky in the sense that neither myself nor my business depend very much on these tools because we do work which is more complex than frontend web apps or whatever people use them for these days. We use them here and there, mainly because google search is such crap these days, but we had been doing very well without them too and could also turn them off. The only reason we still keep them around is that the cost is fairly low. However, I feel like we are missing the bigger picture here. My point is, all of these companies have been constantly hyping a near-AGI experience for the past 3 years at least. As a matter of principle, I refuse to do additional work for them to "make it work". They should have been working already without me thinking about how big their context window is or whatever. Do you ever have to think how your operating system works when you ask it to copy a file or how your phone works when you answer a call? I will leave it to some vibe-coder (what an absurd word) who actually does depend on those tools for their livelihood.

by cma 2y ago

hansmayer 2y ago

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).

cma 2y ago
LLMs are also primarily promoted through the web chat interface, not always magic wonder tools. With any project that will fit in claude/gemini's large context you use those interfaces and dump everything in with something like this: (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 Then drag that into the chat. 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 You can then just hit ctrl+r and type claude to refind it in shell history. Maybe that's too close to "writing scripts" for you but if you are searching a large codebase effectively without AI you are constantly writing stuff like that and now it reads it for you. 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.
- hansmayer 2y ago
  No thank you for obviously good intent on your side, but I am not looking for scripting help here, nor am I a business type who does not code themselves. I just don't want do this when I am already paying for the tooling which should be able to do it themselves, as they already wrap Claude, ChatGPT and whatever other LLMs. And unless you're professionally developing with Microsoft stack, I'd advise to ditch the Windows+MinGW for Linux, or at the very least, a MacBook ;)
  - cma 2y ago
    The tooling you are paying for doesn't work with the full abilities of the context so you need to do something else. Doesn't matter what it's supposed to do or that other people say it does everything for them well, it works a lot better with as much in context as possible on my experience. They do have other tools like RAG though in cursor, and it's much quicker iteration, ultimately a mix of what works best is what you should use, but not just block stuff out out of disappointment with one type of tool.
    - hansmayer 2y ago
      I am lucky in the sense that neither myself nor my business depend very much on these tools because we do work which is more complex than frontend web apps or whatever people use them for these days. We use them here and there, mainly because google search is such crap these days, but we had been doing very well without them too and could also turn them off. The only reason we still keep them around is that the cost is fairly low. However, I feel like we are missing the bigger picture here. My point is, all of these companies have been constantly hyping a near-AGI experience for the past 3 years at least. As a matter of principle, I refuse to do additional work for them to "make it work". They should have been working already without me thinking about how big their context window is or whatever. Do you ever have to think how your operating system works when you ask it to copy a file or how your phone works when you answer a call? I will leave it to some vibe-coder (what an absurd word) who actually does depend on those tools for their livelihood.