This might work only if you have “infinite” compute and infinite tokens.
As someone that used the $20 plan, this pure agentic approach is impossible to do because I’d hit the limit fast and I would end up with less outcome.
What I found that work incredibly well was to provide a human written code as reference, and ask it to extend it. So I scaffold the entire thing, architect it, write few samples code (controllers, services, models, components, database schema, how auth works, etc) so the LLM can have a headstart on their attention (pun intended)
I usually wrote a stub with a lot of details on how to implement it. Something like a higher abstraction pseudo code. Then ask the LLM to implement it.
When it fails, it is often better to undo the whole changes, adjust the stub so it catches what fails before, and try again.
Or, commit the changes, and use a new fresh context and only address what went wrong.
-
Whenever I tried this agentic from scratch approach, I always end up disappointed; both on the outcome and on the limit that I hit before an hour even passed.
Totally understandable. There’s a way to resonate with other outside of music right. Yours was Diablo. I can’t resonate with that. But if you play Wuthering Waves we’d be a great resonators (pun intended).
Maybe the word that I use “exclusively” is too strong. It means I default to listen to my own songs. Maybe 90% of my playlist songs is my own songs. The first song that I listen in the morning is my song. But I also listen to the remaining 10% albeit not often. And that 10% helps me understand other people too ^^
I do share my music with friends, the reason why I can publish my song is because I chose the words carefully, so it might sounds pretty and romantic to others, but to me it might have different meanings.
Is it the right thing? I have no idea mate, I am, still trying to understand - too.
I might misunderstood you or you might misunderstood me.
I was moved by a lot of songs made by artists I never met. But I was moved because of the song, not because of the reason why the artist wrote it. If I can truly understand the emotional state of the artist when they wrote it, I might be able to empathize with them. But that's me empathizing with the person that made the art as a human. Nothing stopping me from doing that as a human, even when their song didn't move me.
I publish my songs under a pseudonyms. They can infer what am I as a person based on the songs that I wrote. They can infer what emotions and feelings that I am experiencing while I write the songs. But it's all inference, unless they know me behind the pseudonyms, they won't be able to relate with me personally, as my real self, not as the songwriter. And I am okay with that.
That's the feelings right? What's wrong with enjoying your creation right?
The song that I wrote has more values because it carries memories, emotions, and my internal state. I'm not saying other songs doesn't have values. It's just harder to resonate with, unless the melodies or lyrics align with my emotional state.
I listen to my own songs because I am songwriter, and still am even when I stopped doing it professionally. I am not doing it for the sake of "I just want to listen to my own songs and I will never listen to others". I listen just because "This song is meaningful for me"
And the "this song" in that quote above, can be mine, or other's.
My point was narrower. When I write the lyrics, choose the melody, and use AI to help turn it into a polished arrangement, the result has personal meaning to me. It reflects my memories, taste, and emotional state. That does not make it superior to other music. It just makes it personal. I can resonates with songs made by other people if it align with my current emotional state.
I just don't spent time actively exploring new songs to listen to. I spent more time writing my own expression than exploring other expressions.
When I said exclusively, its not that I am not exposed to other songs as well. I do follow certain artists that I really enjoy listening to because I find their lyrics and melodies resonates with me, even when its not in the genre that I preferred.
It's just I don't go and explore songs actively. If my playlist suddenly randomize itself (which YouTube Music usually do even when I already selected a specific playlist), I usually just keep it randomizing the songs, I either skip the songs based on the intro or just the title.
Sometimes, I only write the lyrics without any melodies, or just give a base chords for the AI to work with without any melodies, and AI might surprise me on how it suddenly chose a certain melodies or chord progression.
So you can say I'm exploring, but only within the boundaries of the lyrics that I wrote. Or when YouTube Music randomly plays a song for me and I immediately resonates with it.
I did the same too. I listen exclusively to my own songs made with the help of AI.
My styles are orchestra and symphony pop, which I find rare these days. Even if it exists, the lyrics might not be something that I enjoy.
So I just write my own lyrics, decides on the melodies, and put it to AI to create a polished version.
Do I feel emotional when I listen to it? Of course, its my own lyrics that I wrote. Of course I sing along with it because its the melodies I chose.
And its even more emotional because I relate to it.
Someone can create some songs with billion listeners and emotional for others, but if it doesn’t relate to me. What am I supposed to feel?
My listener wont be able to relate with me personally because they don’t know me. But they might be able to resonates with my songs because it triggers specific memories or emotions for them. And for me that’s enough. Let the songs be the one that they resonates with.
The quote you referenced is about “the weird” feelings. Maybe its weird because I didn’t read too many novels. So its totally something personal to me. Not that weird for me = bad writing.
However, I do read a lot of LLM generated output. I spent weeks tinkering with LLM while I was asking it to ghostwrite my novels. So I was exposed to a lot of text that has this weird feelings. Which I eventually felt when I read this article.
Its like hearing a song that has the same chords so many times, and then you listen to another song that had the same chords, you might be able to know that they are kinda the same, even when you don’t listen to a lot of songs.
I tried writing a short novel using Claude Opus 4.6, I gave it outline and raw draft, and the style is very similar to this writing.
I tried to steer it away from this kind of writing because it feels weird. But it always try to output something similar to this. Or maybe I am just not used to reading novel.
So I was curious, what kind of training data was Claude trained on, that its very hard to steer it out from this style.
So I opened my kindle and looking through the recommended popular novels. Just reading through its free samples.
And the similarities are striking. Now, I dont know whether the recommended novel is the training data, or its actually written by LLM. Or maybe its just how novelist writes.
I even tried writing full chapter from scratch. And asked Claude to ghost write the second chapter for me using my writing style. It still wont follow my style and keeps writing in this kind of style from the article.
Not accusing the article of using an LLM to ghost write. Even so its fine to use LLM to ghost write. Its just one anecdote from my side, on how LLM fails to follow my writing style and keeps coming back to its training data.
As someone that used the $20 plan, this pure agentic approach is impossible to do because I’d hit the limit fast and I would end up with less outcome.
What I found that work incredibly well was to provide a human written code as reference, and ask it to extend it. So I scaffold the entire thing, architect it, write few samples code (controllers, services, models, components, database schema, how auth works, etc) so the LLM can have a headstart on their attention (pun intended)
I usually wrote a stub with a lot of details on how to implement it. Something like a higher abstraction pseudo code. Then ask the LLM to implement it.
When it fails, it is often better to undo the whole changes, adjust the stub so it catches what fails before, and try again.
Or, commit the changes, and use a new fresh context and only address what went wrong.
-
Whenever I tried this agentic from scratch approach, I always end up disappointed; both on the outcome and on the limit that I hit before an hour even passed.