I think I'm done using LLMs altogether for coding. I've lost the ability to maintain any kind of flow state, the majority of the time I've spent thrashing on architectural changes that I could have done myself, tests that get manipulated into passing, and having to sift through the magic 8-ball of skills that are intended to get work done (all caps dont do this, please do that). LLMs appear to be fruitful as essentially a research search engine but I'm pretty much done with them for coding. This has been an enormously expensive waste of time and to add to it a general atrophy of skill.
I've found a middle ground where I sketch architecture by myself or using AI as an advisor. Document all my decisions and key points and then use AI coding to extend foundation and add features.
Basically what I do:
- I have an idea of some architecture solution for the current task
- I try to sketch it myself and use AI to find some flaws I missing or try to find alternative approaches
- Review comments from AI and summarize into final decision
- Implement foundational layer myself or using AI to help me
- Document all crucial points like module boundaries, dependencies, key strutural elements
- Use AI to add features on top of newly built architecture
Man, I just got into a killer flow with it, minimal guardrails, having it doing lots of independent work all the way to task completion, and my company dropped our prem token allowance to 60%. I was hitting 2k+%. So I just got hamstringed. Corporate pulled a complete 180 on us.
Would love to hear if other are in this situation.
I would describe the flow-state challenge as: Needing a way to make the agentic coding process continuous rather than intermittent. It used to be that I could let a semi-visual map of the code build in my head, then work based on that map (continuously) for a few days. Now I have to basically load a whole new map into my brain for one task, write a detailed prompt, hit enter, then dump all that "context" to go work on a prompt for some other part of the code. I have been experimenting with tools and processes that let me reduce the amount of context switching between each task so I can keep more of my mind at a higher "architectural" level, which feels more continuous.
A helpful middle ground I’ve found is to build out the architecture you want, but stub out the tedious function implementations you don’t want to do yourself.
And by stub out I mean write the function signature yourself, including parameters it’ll accept and return types. Add a comment if necessary about what it will do.
I think using skills like grill_me is supposed to make the experience a lot more user-involved and fun, rather than just being an accept_permissions driven process and the feeling like an observer.
This is actually a great idea. I build a whole thing into my agent management platform that basically treats tickets (kinda like Jira, but not) as a store of shared context, then lets me launch prompts into the terminal with all the ticket history as context (all the file changes and prompt history, etc). I find that just having that context improves consistency quite a bit without pulling in a whole chat.
Share your POV about the flow state. I lost that too, but found myself a few days ago sitting in front of Claude prompting until 5am in the morning. The flow is different, less satisfactory for sure than writing code yourself.
Main reason I am sticking to using AI is convenience, speed and the perception that the job market expects you to.
bro, i cannot explain that feeling, i recently building a project at first ai done a good job but then i try to change something and it make the whole project trash, still stuggling to make it done, I'm done using AI too
I don’t think the problem is AI, the problem is the way we use AI and what we ask it to do. Flow comes from having a clear mental model, making decisions. If the LLMs model is making most of the decisions, I’m left trying to reconstruct someone else’s thinking after the fact.
I think that’s the key. Don’t hand over the thinking, use it as part of the loop. The more I use it, the better I get at knowing what to ask for, what to ignore, and when to take back control.
think simple. you are human, ai is a car. you can walk but if you ride a car you can arrive fast. you don't need to forget how to walk while you riding a car.
another side i think for myself; if you are developer you are somewhat right. if you are maker, no problem you can use ai as much as possible. bcz i know i can write completely if the ai is turned off.
If you ride a car, you can arrive fast. Maybe. If you don't wreck the car. Or get hit by a bus. Or have your tire go flat. And your engine doesn't blow. And you don't get pulled over by the cops. And there is no construction on the roads. No accidents. No bridges washed out.
So yeah, your analogy works. Better than you intended.
Basically what I do:
- I have an idea of some architecture solution for the current task
- I try to sketch it myself and use AI to find some flaws I missing or try to find alternative approaches
- Review comments from AI and summarize into final decision
- Implement foundational layer myself or using AI to help me
- Document all crucial points like module boundaries, dependencies, key strutural elements
- Use AI to add features on top of newly built architecture
Would love to hear if other are in this situation.
Curious what others are doing to solve this.
And by stub out I mean write the function signature yourself, including parameters it’ll accept and return types. Add a comment if necessary about what it will do.
https://github.com/mattpocock/skills/blob/main/skills/produc...
Main reason I am sticking to using AI is convenience, speed and the perception that the job market expects you to.
another side i think for myself; if you are developer you are somewhat right. if you are maker, no problem you can use ai as much as possible. bcz i know i can write completely if the ai is turned off.
So yeah, your analogy works. Better than you intended.