Thread by @rvm0n_

how to actually vibe code a complex project, as someone that doesn't know how to write a single line of code

these are the 5 most important rules you should follow:

1)
prompt. prompt. prompt.

i cant emphasize this enough, if you don't know whats going on in the code, you need to spend a good amount of time to describe your feature requests, issues or questions clearly and detailed.

also try to add images whenever relevant. i also like using a marker/draw tool on top of it to make it even more clear

2) console.log is your best friend

if you're stuck on the 3rd iteration of fixing a bug / implementing a feature and you can't get it to work then console.log it!

it takes like 1 minute to implement various logs and the agent generally will know what he needs, through this, you can actually get down to the core of the issue without being blind.

remember to delete the logs once you fixed your issue

3) .md files are beautiful

these are for bridging the gap between your understanding and the ai agent's understanding

my projects usually consist of

- core-idea.md - this is where you describe the general project/product you want to build. WRITE THIS IN YOUR OWN WORDS

- various architecture .md files - these are depending on your complexity, for example, i have a .md file that counts as a source of truth for midi logic, for canvas rendering, for sound and so on.

- bug/feature .md files - if you cant fix a bug in a session and you are tired, ask the agent to create a .md file and write down the initial problem and all the things he tried so far. try again in the next day and you can usually fix it

4) actually understand the fundamentals, ASK STUPID QUESTIONS

a LLM doesnt judge you, ask the most basic questions about how something works, why something is handled in this way, why this is used over that, how other apps usually do it

this will steer you into the right direction from the beginning and you will actually learn a ton

5) use AI everywhere

you want to implement a specific thing but have no idea how to and the ai agent can't add it correctly?
-> do a deep research on how competitors do it (i had this with my infinite piano roll canvas for example)

you added logging and took 10 iterations but still cant fix the bug?
-> let gpt-5-codex, claude-code, gemini and what else do their own analysis, feed them each others reports and get down to the core issue

these are the most important rules when it comes to actually getting shit done without knowing what you're doing

by working on different projects you will get a very good feeling of when something feels right or wrong and you will learn a ton along the way.

i hope this helps!