Discussion of AI tools for development

[hjtrbo]

Edit: Forked from viewtopic.php?p=138011#p138011

I'm just starting to dive into this "vibe coding" deal as a curiosity experiment. I have a couple of questions related to Claude code...

How have you gone with your plan limits?
Which plan are you on?
Are you utilising CLAUDE.md, rules & commands files in addition to your skills file(s)?
Do you reference your skills file at the start of every chat session?

Claude user account settings ->
Capability-> Have you turned on "Search and reference chats" and or "Generate memory from chat history"
Usage-> Is "Extra usage" turned off?

[antus]

I jumped in on claude because everyone says its the best but I've found its quite dumb, goes around in circles and blows the limits easily without reaching the conclusion without a lot of configuration and hand holding. And I plan to cancel my Claude account and go back to chatgpt with its codex plugin in vscode at the end of the month. I find it smarter, quicker and a lot more to point.

[hjtrbo]

I have found exactly the same thing as you antus. I frequently blow my 5 hour limit in under an hour (Pro plan). The main issue for my project is I have esp32 build code and lvgl framework code in the projects root folder. It was always searching through code it shouldn't have been and second guessing itself. That said, I have been building up CLAUDE.md that now instructs claude to never go into those folders unless I let it. It must reference it's own internal knowledge of the lvgl api before it's allowed to go searching through the framework code by itself. I'm also compacting frequently when reaching mile-stones. If I feel we've had a good session and made some key lvgl api discoveries worth keeping, I update or create a new skills file. These are getting me longer sessions but I'm still hitting the limit. Its the reasoning or "thinking" that it does seems extreme. My opinion is that out of the box claude is hungry for tokens. I've only being using it for 2 days - I'm forming the view that it will require significant user intervention to get more efficient use out of it.

I think that is the new skill (or art if you prefer) to learn - controlling the AI efficiently as opposed to writing code by hand.

I'm open to giving codex a go.

[antus]

Get a trial account of chatgpt and install the offical codex plugin to vs code and open your project as a folder (I use web login not api token, its easier, though you could use API if you wanted). It'll work. I usually use codex-5.3 in medium mode because its lighter on tokens. But they merged codex the AI in to the mainline to chatgpt 5.4 or 5.5 (when it appears there) so they should be better, but in general 5.3-codex in medium thinking seems to deliver at light weight out of the box. You can also turn the thinking level up if you are giving it a complex problem. But a complex ask on high or very high likely will use all your tokens in one or two hits.

On the other hand qwen 3.6 has come out. They claim its frontier model capability, run locally. But its very slow on even a pretty good CPU from a couple of years ago. I think in my case DDR4 is the bottle neck. DDR5 machines help. Macbooks with unified memory are better again. Or Intel server gear with 8 channels of DDR5 (but prices are insane, even second hand. They're probably the ones the cloud providers use).

You can quantize models down in size, but you really have to neuter it to fit it on a 16Gb GPU like mine. I'm looking at versions like https://huggingface.co/mradermacher/qwe ... ed-i1-GGUF running in LM Studio behind cline for vs code. But mixed results so far. Any recommendations for models like that locally on a 16Gb GPU I am all ears.

If I was buying hardware I'd go a current gen mac mini with the lower CPU and the max 32Gb of unified memory for about $1500. Not as fast as a GPU, but probably best bang for buck for that kind of price. Or an old RTX 3090 25Gb GPU.

[kidturbo]

After two weeks running both Claude Code in VScode, and GPT in Codex, sticking with Claude on the C and Python work. It's logic is just faster, and seems to pick up where I left off 80% of the time.

Now I put GPT on some CANbus reversing and DBC creating duties, after Claude started to struggle with analyst duties. GPT pick right up on the "Other Agents Work" tagged in folders and files, and took a differ approach or logic. While I like GPT's approach to reversing, it often goes in to an "Ethics" review of what it's being ask to accomplish.. With Claude, I had to write up a chat on ethics.. Both models seem decent at reading PDF files, or printed data saved as PDF. That's helpful for a lot of work.

So running both agents on the same projects or folders seems to make each act smarter. While I continually have to aks GPT create md files everything or it gets lost, it does really good at converting that data to Human Useful html. Then I just refer each agent to said html files as our main Status or Reference docs.. Seems to speed up the thinking if noting else.

Attached screenshot from working html page with list of sorted msg ID's, that can jump to, and actually prints out on paper is a usable book layout..

[kidturbo]

Edit: Forked from viewtopic.php?p=138011#p138011

I'm just starting to dive into this "vibe coding" deal as a curiosity experiment. I have a couple of questions related to Claude code...

How have you gone with your plan limits?
Which plan are you on?
Are you utilising CLAUDE.md, rules & commands files in addition to your skills file(s)?
Do you reference your skills file at the start of every chat session?

Claude user account settings ->
Capability-> Have you turned on "Search and reference chats" and or "Generate memory from chat history"
Usage-> Is "Extra usage" turned off?

I run run Claude Max, while running inside VScode, is about all I can use per day. I hit a 5hr limit maybe once a week, if on it for 8hrs straight. The lower plan was limiting me out every couple hours.

I make sure it is updating claude.md every session. Along with creating .md files on major topics. Then keep a changelog. As for skills file, set up once with rules, and general preferences is all I did. The claude.md seems to carry that over.. Like what to call you.. Every session greets me as Kidurbo, and I ask it to show me where we left off. Then it's typically not asking me thing twice. If it does, or I catch it missed something, point it back to the .md files to explain Why..

Capability, yes to about all. Max Reasoning, takes more tokens, but better results if complicated problems. Extra Usage to OFF.. It ask me to turn that on when I started running on Claude Desktop and Phone linked. As uses way more tokens than in VScode.

I suggest watching some youtube videos on best commands and prompt setups.. Then, Ask it "what would you like to know about me to help understand my goals." It will walk you thru getting everything set and happy. Then write it a couple paragraph prompt of something you want to build, analyze, ect. Tell it to Ask You any questions about tasks, and then hit send. Once you get comfortable feeding it good prompts, hit Automode and walk away for a while..

[hjtrbo]

I upgraded to Max 5x. That suits my needs better whilst I'm learning. Finally got through consecutive 5 hour windows today without even hitting 50%.

Probably the biggest win I've had is to put it into "plan mode" and tell it to ask me lots of questions before any refactor or new / changed feature. And like you said kidturbo, plenty of claude.md updates.

With respect to this vibe coding thing, I think if you have a specific project in mind - pay the $160AUD and get her done within a month. Then downgrade to the $34AUD pro plan for general casual usage. If writing code 'the old fashioned way' I can see the pro plan being plenty big enough.

There is a downside to vibe coding, you lose all familiarity with the code base. And the code the AI produces is nothing like a normal non-professional code hack writes. e.g. the AI variable will be "kb_obj_en" where ol mate here would write "enable_onscreen_keyboard".

[antus]

Ive noticed it will copy style. If you have a well architected, consistant large app. Sometimes ask it if it would consider refactoring the code base, or ask it about app architecture improvements, expecially when your starting out helps. If you just let it go and the app gets messy with if/then/else statements all over it'll keep doing that. If you read the code, and tell it to make and use libraries for something, you can get it to work much better. If you have a large traditionally written app with not much tech debt and a clean design and you watch what its doing you can keep getting pretty good results. I think when vibe codeing the humans interaction with the AI and focus on quality control, rather than "it works, ship it"[tm] is what makes one coder better than another and stops (or reduces) you shooting yourself in the foot later.

[Tazzi]

I find AI topics to be so controversial.
One half will praise it like a god.
Other half will bag it out, saying how useless it is.

I honestly see it like a hammer.
A carpenter with 30years experience is going to be able to use the hammer very well.
The apprentice still has alot of sore thumbs and bent nails in his future while learning to use it.

AI is no different, anyone expecting it to make complex things from scratch with minimal effort or providing their own previous 'work', and not learning themselves are probably not understanding its purpose.
Projects get out of hand pretty fast as they get larger, I am starting to see a rise in people asking help in debugging vibe coded applications when AI goes around in circles on a problem. If you don't understand your own code base, then it becomes pretty damn hard to debug it.

Now I am being bias because of how I use it, but its simply an extension of the work I already do, it speeds up my work flow rather then repacing it.

[kidturbo]

Agree if you you feed it examples of coding styles, it tends to follow that pretty decent. If ya tell it to just build from scratch, it seem have it's own way of doing things. I started using claude to debug and clean up a decade old project folder full of variants on same hardware. It separated new code sections into nice includes, cleaned up all mine, and rewrote it's own comments like a pro. Can move between projects now, and applies what and how we did things on the previous ones.

Then I turned it loose on this Vibe Coding approach to see why all the rave. And while it can create a working app or game from scratch quickly, it seem to dump 80% of the code in a single Main file, with hardly any comments explaining it. Guess it doesn't need any explanations taking up space..

Also forgot to mention that Planning mode difference above. Also found it worked best at helping Claude understand what you are trying to accomplish. I had been using planning mode before each change, or at new chat window prior to the Auto Mode release. So It already had the saved session chats from earlier, and I was just thrill to not need to Okay every change. Not the same if you just feed it a single prompt and set it free.. I find myself watching it's thinking more now, and having to put back on track. That wasn't as noticeable when using Plan mode first.

Have also found you can run multiple chats in same project now, and it will recognize the other agent has changed code. But I still make wait until one finished, before letting second chat make any hard changes.. So will run one doing research and write .md files, and other chat is changing code based off those same memory files.