Claude Code research preview - the future is now!
 

this thing is awesome! it's terminal-based, it creates/edits/deletes files and folders, installs and runs software etc.(what could go wrong, right? :D) and it's smart. I've been using it since yesterday and am impressed.

it's way more expensive than chatgpt / cursor and others. chatgpts plus monthly subscription fee easily gets spent in one hour with claude code. and it's worth it, it's so damn good.

I use Claude 3.7 through GitHub Copilot agent mode and it's also quite nice, even though I still prefer Cursor for its dev experience.

Now I also think that I need to try Claude Code...

How do you use it? Is it like small regular dev tasks like "implement this feature" or sth, or you create an app from scratch?

Also, do you really like the terminal-based control?
I used Aider & Gemini before and I would still prefer Copilot, because it's just much more clear and UI friendly

I've been using 3.7 since it came out pretty heavily. first in cursor, now in claude code. even though it's the same model, claude code doesn't mess with context size to make it more cost efficient. you yourself decide if and when its time to create a summary or clear context all together. so in claude code 3.7 does actually go through as much of your codebase as it deems necessary to complete a task. it's so much better, and it burns through tokens like mad :1orglaugh

how do i use it? well, first i tested it to help me implement a proper authentication logic(quite involved) into the app i've been working on. no, it didn't do the whole thing correctly in one go. but still impressed me. since then i've completely switched to claude code. it's a research preview and they're pushing out updates often.

i too was skeptical about the terminal thing. but it's great, anthropic have done a great job with this thing. just try it, man - you will be impressed.

speaking of Anthropic, that is a cool company. started by siblings after they left openai, then few more people left openai for anthropic. SBF invested $500m, then amazon a few billion. their claude 3.5 sonnet has been dominating ai-assisted programming since it's release just over a year ago. and now they release this magical thing.

here's the Lex Fridman interview with the siblings

https://open.spotify.com/episode/69V...d7a9a8903d488a

:thumbsup

What do you do specifically with the scripts, files or code you generate from it?

Yea, this one seems wild.

I tried it out based on your recommendation.

2 dollars and it coded a complete, pretty much bug free entire checkout for a system Im working on.

it used all my existing models, code style, css stling, without any prompt

( i actually pasted the files I WANTED to prompt it for - and it accidentally submit )

it did exactly what I wanted with no prompt just based on the fuckin file names - wow

I will bro, now you've convinced me to 100% try it.
Thanks!

debug, deploy, make millions(hopefully)!

I guess I am asking, "debug, deploy, and make millions ..." with what? What kind of site?

at the risk of sounding like sleazydream... it's a novel idea and I'd rather keep it to myself until the release :)

No, I wasn't asking about what your project is. I don't ask people about their own stuff.

I just meant what are most people doing with the scripts they generate with AI to make money? What kind of sites are they generally building?

I see lots of store sites that sell tons of aromatherapy or nutraceutical products claiming to make $40k month. But for some reason they are selling them at 1.2 million. I guess they have enough money and want to 'help' others lol

for those of you who don’t know about mcp servers…

so, claude code by itself only has access to the contents of the folder that you run it from. but there is this thing called model context protocol(also invented by anthropic btw). using that you can connect it to databases, the file system, the browser and much more. this is limitless!

the claude crawler bot is very intrusive on your website if you do not have it blocked in robots.txt, it will send over 50,000 hits in a few hours

I have blocked them all now

Https://mcp.so

Basically giving your ai agent superpower 😉

Can you expand on this?

I'm making a system to create fully automated video compilations and I've been using ChatGPT (sucks now) Grok (sucks less) and Claude (fucking rockstar so far), but each of the 50 different things my complex system does each requires its own .bat file and ods files and API calls to online AIs to create the finished product... which is not very polished yet. It's getting better, but the compilations are not usable. Close, but no cigar.

Super interested in this thread as ChatGPT seems to have gone backwards. It actually feels like it is worse now than last year.

There's also aider via ollama they getting better every update as well and its via terminal. Oh and open source!

???

Aider? Ollama? What are these?! :upsidedow

https://www.youtube.com/watch?v=AJpK3YTTKZ4

the problem you will face is that you can't afford to run it.

Its expensive.
That said I'm taking a hybrid approach and mostly using cursor, while having claude code installed and funded, when I run into a real headscratcher.

If you are trying to code using the chat interfaces and pasting back and forth, that's just not the way.

Even if you do get what you want, the code will be horrific

Im on OSX so never faced any of the bat file problems you're having cant help ya there.

How much does it cost?

Stringing together variuous .bat files is working for me for the most part. I'm 90% there when it comes to the compilation building, they're just not polished enough to post yet.

Runs about the same as a webcam girl, a few bucks a minute :1orglaugh:1orglaugh

Ive never touched a bat file in my life, so no idea what you got going on there :upsidedow

Ollama lets you run AI models locally on your computer, while Aider is an AI coding assistant that helps with programming like claude. 🫡✊🏾

You code and have never touched a batch file?!?!?!?!!?

correct :)

thanks for the lead on this, somehow I missed it.

Im using cursor for light liftting and then have claude code installed and funded, ready for heavy lifting.

I burned through a few bucks hella fast, so I just have it ready now for real head scratchers... cursor does fine in most cases, no need to light cash on fire

so a few days later, I'd actually like to ask why you think Claude code is worth the cost

While initially impressed it did highlight a flaw in the way I've been using cursor. ( giving files as context actually narrows the context, rather than amplifying it )

since then I've not needed to invoke Claude code once.

I still like having it available, but not sure why you think its worth running as your main agent? I don't see it, but could be missing something.

Claude is fucking amazing and I'm not even using what you're using!

Wish I had $28 for this... fuuuuuuuck. Down to my last 4 eggs, 2 hamburgers, and a single bun until next affiliate payment! :helpme

https://i.ibb.co/DH60WX3L/fuck.png

Does this cost like $1 per question you ask for help with?

Holy SHIT I've been using Claude free now and it seems soooooooooooo much better than ChatGPT free!!!

3 days of Claude and the project I've been working on with ChatGPT, unsuccessfully, for 3 weeks is ALMOST ready. The final 2 issues will get solved today, I bet.

...unfortunately a yearly plugin renewed overnight, hitting my overdraft, and causing me a $-45 NSF fee hit me, which takes a big chunk out of my next affiliate payment.

Being poor actually makes you poorer when this stuff happens, because that's $45 more the bank is taking from me just because I don't have a float to survive these things with. Billion dollar companies profiting because I struggle to survive! Holy fuck, well, today I am going through PayPal and cancelling all automatic payments I may have forgot about, all while having 0 food to eat. Good times, good times...

At least Claude is fucking incredible, it is SO much better than ChatGPT so far. ChatGPT, at least the free version, is hot garbage at the moment. Claude gets things right in 2-4 prompts and ChatGPT, after like 20+, just hallucinates and starts sending me Python code randomly and I have to remind it what tast we're even on.

https://i.ibb.co/jk84vNy2/gettingoorer.png
https://i.ibb.co/Z1RXj4M8/money.png

Things are moving really fast these days, at some point you just have to accept what is and go build shit. :1orglaugh

Have you tried the cursor trial?

You are using the LLM model and chat interface, NOT their model through an agent.

Totally different experience.

Nah, not some of these guys :1orglaugh:1orglaugh

I'm codin' with Claude right meow, but I keep hitting the timer and need to fuck around for a couple of hours on GFY and other places while I wait for the free timer to reset... :1orglaugh:1orglaugh

I showed you how to start projects and it cuts the timeouts way down. Having two accounts that share project files let's you keep going, most of the time but not always.

its too easy right now to get caught in an AI loop and accomplish nothing.

Im always buidling, making the most of any and every tool possible.

We can refactor later!

I spell the goal out clearly in the project overview... and have to requently remind Claude to remember the goal or else he's trying to get me to include all kinds of shit and sidetrack the whole damn thing :1orglaugh

I wish I had more pre-AI coding experience like some of you guys, but Claude is still crushing what I used to spend thousands trying unsuccessfully to get developers to do.

You also are using chat interface?

Get away from that man, its a trap! cursor is the way right now.. IMO :2 cents:

Thanks for talking about this bro, downloading now! :thumbsup

https://i.ibb.co/Gv7pQGXX/dl.png

I have it installed, plan on trying it after this project is done. So maybe tuesday or wednesday.

I don’t think you’re missing anything, I wasn’t using cursor to its full potential. I used to add files into the chat as context and prompt it, and it seemed amazing especially considering that before then I stayed loyal to sublime text for 10+ years.

yesterday I tried cursors agent feature and it’s a whole new level. the new version is very nice as well. so now I’m using both cursor and claude code, trying to spot differences in logic / code quality :)

I was looking for a CRM for managing my script business and could not find anything that looked suitable - I asked Cine to make me one and it spat it out 3 minutes later - Amazing!..

Asshole wants to take over the world on day 1. Takes bm a month to debug a simple WordPress plugin circa 2008, I'll settle for 4 days to world domination.

Yea, thats what I discovered too, I thought adding the files would be better it narrows the context window.

Sublimetext brings me back... way back!

I would definitely look into coding tutorials so you can at least supervise it properly, as it will certainly write some shit code.

It will aslo "fake" stuff, so it looks like its working... going back to the tax class I mentioned.
It clearly used provincial rates for canada, but hard coded, instead using my tax class..

So if I didnt know how to code, and was only looking from the client side, Id have not have caught this at all...

be vigilant.

Also use other AI to check its work... where you can

if you start using cursor, dont forget to use mcp servers, for example:

Sequential Thinking MCP Server
An MCP server implementation that provides a tool for dynamic and reflective problem-solving through a structured thinking process.

to install on cursor
npx -y @modelcontextprotocol/server-sequential-thinking

Keep building guys!!

That's really interesting with the tax thing.

Do you have any recommendations for code tutorials or online classes?

My YouTube always wants to teach me how to use AI, but I get AI, I love AI. It's knowing how to, like you said, supervise specific types of work.

Had a lot of success coding WordPress stuff with it because I have some basic ideas of how that code works, but I'm out of luck on more sort of what I think of as functional code.

I really couldnt tell ya where to begin learning code in 2025, I started in 98 so thats worlds away, just evolved with the internet I guess.

I can't stand youtube tutorials.. too slow but

When starting with a new stack I typically load up their docs, then google "stack name best practices"

From there, for you I'd google "whatever stack beginner tutorials"

if you're using chat interfaces, use a new chat window, or even another AI to explain in detail what the code is doing ( modern rubber duckie coding )

this should allow you to confirm that the code does indeed do what you think its doing.

Different chat windows are key there.. as the same window will happily lie to you and confirm its own bullshit code :1orglaugh

You can simply ask ChatGPT to teach you. In fact, prompt it like this:

https://i.ibb.co/KzG7SHL2/promptpirate.jpg

Thanks. I've definitely had AI lie to me. ChatGPT has confidently repeatedly sworn to me it put the data points in the correct part of the spreadsheet, when it patently had not. Grok is the worst with this (for me so far YMMV) lying about its abilities and swearing links go to things they do not etc. Claude is my friend though and is straight with me.

I've had Genisis (Google) tell me I was wasting it's time because I didn't like an edit it made - We had a full-blown fight!..

:Oh crap