Trying AI programming
Discussion in Programming
[@mmHg](/profile.html?user=47) I want to try AI programming. I have macOS 10 . I don't think Codex or Claude Caude run on this OS. I don't want to update my OS. When I get back to America, I could use another computer to try this. For now, Claude recommends I use Aider. What do you think?
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=114#p114)
>
> [@mmHg](/profile.html?user=47) I want to try AI programming. I have macOS 10 . I don't think Codex or Claude Caude run on this OS. I don't want to update my OS. When I get back to America, I could use another computer to try this. For now, Claude recommends I use Aider. What do you think?
You could get a subscription for OpenAI or Anthropic and then open ChatGPT/Claude in your browser -> select ChatGPT’s Work Mode, create a project folder -> add the instructions + context files + library files that you want -> start a chat where you upload your src file and give it instructions. It should spit back a new src file for you.
The main advantage of using ChatGPT Codex over using ChatGPT’s Browser + Work Mode is that everything is the same, but Codex automatically modifies files on your computer for you and also uses the terminal for you.
I don’t know what Aider is. It looks interesting. It appears to be a harness, similar to oh my pi. I heard good things (from a friend of mine that works with macOS and Qwen) about **pi** / **oh my pi** but I never tried it. Supposedly, it lets you use different AI models for different tasks concurrently.
What are you using exactly? Apparently these browser tools all use GitHub, so they are useless.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=139#p139)
>
> What are you using exactly? Apparently these browser tools all use GitHub, so they are useless.
I use Codex, which is just Work Mode done locally instead of on browser.
You could use ChatGPT’s Work Mode on browser. Alternatively, you could use Claude on browser and create a project folder. Neither of these require GitHub.
> [mmHg](/thread.html?thread=113&post=142#p142)
>
> Alternatively, you could use Claude on browser and create a project folder. Neither of these require GitHub.
I tried Claude on browser and it seems to require GitHub. Please try it yourself. Here is a week pass:
https://claude.ai/referral/GufFFNT-1Q?s=claude_ai
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=160#p160)
>
> I tried Claude on browser and it seems to require GitHub. Please try it yourself. Here is a week pass:
>
> https://claude.ai/referral/GufFFNT-1Q?s=claude_ai
I tried it and it doesn't require me to use GitHub. It can simply feed me back modified/updated src files that I can download and put anywhere on my computer. Honestly, even the free plan for Claude should be able to do what I described.
> [mmHg](/thread.html?thread=113&post=170#p170)
>
> I tried it and it doesn't require me to use GitHub. It can simply feed me back modified/updated src files that I can download and put anywhere on my computer. Honestly, even the free plan for Claude should be able to do what I described.
This seems very clumsy. I mean could you effectively work on Textile like this?
I will try Aider when I have time. It looks like the most sane option.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=173#p173)
>
> This seems very clumsy. I mean could you effectively work on Textile like this?
Yes, you just have to create a folder/archive for all of your src files and keep it well organized.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=179#p179)
>
> I will try Aider when I have time. It looks like the most sane option.
Sounds like a plan.
On a little sidenote, I am going to experiment with codegraph, hoping to reduce token burning when I work with Codex.
> [mmHg](/thread.html?thread=113&post=184#p184)
>
> On a little sidenote, I am going to experiment with codegraph, hoping to reduce token burning when I work with Codex.
I predict that this will no difference with Luan because it is trivial to figure out where things are in Luan. If it does make a difference, then AI doesn't understand how Luan is organized.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=190#p190)
>
> I predict that this will no difference with Luan because it is trivial to figure out where things are in Luan. If it does make a difference, then AI doesn't understand how Luan is organized.
More testing is needed but I am excited to say that after training codegraph to understand Luan, my token burning indeed reduces.
I think the reason for this is that it already burns tokens by using my context files that are required in order for it to understand Luan anyway, and said context files were not pre-indexed until I started using codegraph.
I tried Aider. It is useless because it can't test its own code. Claude in the regular public website can run and test code, and can read from a Mercurial directory, so I will play with this. I see no benefit of Claude Code on the web over just using the generic Claude web interface.
> [fschmidt](/thread.html?thread=113&post=237#p237)
>
> I tried Aider. It is useless because it can't test its own code. Claude in the regular public website can run and test code, and can read from a Mercurial directory, so I will play with this. I see no benefit of Claude Code on the web over just using the generic Claude web interface.
This worked very well.
https://claude.ai/share/7326dd28-8269-4d14-ba32-2a1e5ef0f8d6
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=242#p242)
>
> This worked very well.
> https://claude.ai/share/7326dd28-8269-4d14-ba32-2a1e5ef0f8d6
I also see no benefit of Claude Code, so I chose Codex from ChatGPT. ;)
Yes, you could achieve a perfectly satisfactory result using Claude (or ChatGPT) from the browser for software development. However, aside from the browser being unable to modify your local computer files for you, you also won't be able to use codegraph or potentially other AI tools or plugins.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=242#p242)
> https://claude.ai/share/7326dd28-8269-4d14-ba32-2a1e5ef0f8d6
I read the second-to-last post that you wrote in your chat with Claude, and you mention using a separate chat for Luan source, reference docs, etc. This is completely unnecessary if I am understanding what you're trying to do. What you want to do is create a "Project" using Claude from browser. Look for the Project tab on the left side of the browser. This will give you the ability to add context files, instructions, etc. that apply across all chats and future chats within the project folder itself.
> [mmHg](/thread.html?thread=113&post=250#p250)
>
> I read the second-to-last post that you wrote in your chat with Claude, and you mention using a separate chat for Luan source, reference docs, etc. This is completely unnecessary if I am understanding what you're trying to do. What you want to do is create a "Project" using Claude from browser. Look for the Project tab on the left side of the browser. This will give you the ability to add context files, instructions, etc. that apply across all chats and future chats within the project folder itself.
I don't think you understand what I will do. I fundamentally need 2 different kinds of chats.
A Project sounds like a bundle of stuff. Claude said to use a "Skill" for adding files. I don't know what the difference is, but I will ask Claude when I get to it.
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=261#p261)
>
> I don't think you understand what I will do. I fundamentally need 2 different kinds of chats.
>
> A Project sounds like a bundle of stuff. Claude said to use a "Skill" for adding files. I don't know what the difference is, but I will ask Claude when I get to it.
In retrospect, my understanding is that you want to integrate Luan with AI since programming without AI will die.
Also, let me know how it goes whether you use a Project folder or a skill!
mmHg
edited
The way that Codex uses your computer for you is it takes a screenshot of your screen display, interprets the screenshot, decides what to do, and then clicks for you. This is perfectly adequate for navigating webpages, but interpreting every screenshot is not viable for any type of reinforcement learning.
I asked Claude for a reference to my approach. Here it is: https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices#develop-skills-iteratively-with-claude
mmHg
edited
> [fschmidt](/thread.html?thread=113&post=392#p392)
>
> I asked Claude for a reference to my approach. Here it is: https://platform.claude.com/docs/en/agents-and-tools/agent-skills/best-practices#develop-skills-iteratively-with-claude
Looks like a great article. Some of the information can be used for other things besides creating “Claude Skills”.
It seems like Skills and Projects should both be used. Skills appear to be something that you can reuse across multiple projects. In your case, you can create a skill and then create a project for the Luan source itself that can then be used to fine-tune the skill that you created or to create separate specialized skills for future projects. Then, you can create Project folders for projects using Luan and reuse those skills.
[This](https://developers.openai.com/blog/rethinking-skills-and-prompts-for-gpt-6-astra) looks related.
Log in to reply.