Someone said “Sometimes less is more” - I’m amazed by some numbers here, seems like i need to boost up my productivity. But i do ‘Shift Right Testing’, each edit i immediately test in production, as it’s my own SaaS and it must be perfect.
I’ve only been a user this month, and have been working on a Blender Add-on. I seem to be getting infinitely closer to the end result I am looking for, but Claude 4 Sonnet seems to really want to break everything whenever a new issue occurs. You’d think a Python based add-on would be relatively simple for this AI to figure out
Everyone: I did 100k lines, I did 500k lines, I did 800k lines
Me: Do I really want to use another request??? Nah, I’ll just do it manually and save it for some other day… (Funniest part is there wasn’t any “some other day”)
Nah nah nah, change your username buddy, it should be godcode, not cocode… lol, you are the king I guess haha. meanwhile me <50K in last 3 months.
If you dont mind asking, what industry is your application?
Just a couple lines over here…
I use Claude 4 sonnet thinking for all my coding now. Was using claude sonnet 3.7 and gpt 4.1 max last month.
I create all my prompts using home websites of Claude, openAI, and Gemini Pro. (I pay for all of them). I just leverage their strengths and use them synergistically to help me research, come up with ideas, fine tune prompts, organize plans, phases, etc. (4.5, 4.1, Opus, 2.5 Pro DR, Regular) then feed all of it to Claude 4 Sonnet in Cursor AI.
I put in about 20 hours a day on average.
One piece of advice I wish I gave myself in the beginning is to have a master markdown file, and with every prompt I now include a master set of instructions and guidelines. When a project gets this big you need it to constantly do deep audits to avoid duplication, consolidate, unify, optimize, etc.
how the heck is everyone else able to use this tool seemingly more effectively than I can? I have an Instructions.md file the tells Claude what to do and what not to do yet it’ll fix 1 thing then break 3 other things. I’ve also got several files laid out that handle different chunks of code, and yet i am still just going in circles! D:
and what application is this guy working on for over 3million lines to be written and changed?
Some of you guys are monsters 0_0
the bump from the 8th is me working on litechat it is a static (html/js/css) local first chat app…
With these I added MCP support (with a http to stdio proxy), prompt library, a race mode, mermaidjs support, AI message edition and git sync for the conversation…
I can barely keep up with how fast my codebase is changing… how are you guys managing, or are you just going full “vibe” ?
Overwhelming 
How many hours do you work? 25 hours a day?
yikes, and its 45 days lmao,
You can try to overcome some of the thinking models tendency to go wild with python scripts in some cases. F.e. when it was ■■■■■■■ up my json structures, forgetting keys, naming them wrongly etc. I used a python script to spit out a list of the mistakes made, running live to adjust and cross of tasks of the list to save Claude from editing the task file (state management) but give ability to come back to a true task list outside of decision of claude (f.e. it didnt find most places where keys were crooked). As for web development, I found it can be extremely powerful to have customizable testing frameworks and spending a lot of time on perfecting them. In one project with 600+ files I use vitest, vite, jest. It generates 1:1 mock files, maintains perfect sync. This enables me as a no coder to bridge many gaps, my own gap, the AI stupidity-genius gap and more. Input on how I can supercharge more always appreciated
I put forth my humble entry for the p*ssing contest. A lot of that is probably Claude 4 Opus usage for complex document parsing that required trying different LLMs (in the end it was Gemini Pro 2.5 that came up with the solution though).
I’d like to believe my prompting and contextual management has improved significantly in the last month, although there’s still a lot of trial and error; I’d guess about 50% of those calls generated effective results.
