I’m struggling to find a good balance between the usage of AI and working in the Engine. I’m trying to use to implement some of the parts of the game, and I’m trying to focus on the:
- Request a plan of action for feature
- Review the plan of action
- Ask for the execution of the plan
But the plan is already broken, dozens of issues in the planning already. I can see some good ideas, some good principles, but a deeper look and I can see the flaws in the plan. Should I invest more time the planning? Or invest more time refining on the implementation of the plan?
From those who use this type of tooling, what is the workflow that gives the 100x productivity boost that is claimed.
From those who use this type of tooling, what is the workflow that gives the 100x productivity boost that is claimed.
You don’t. As your own testing has shown that number is a lie
I know that on the empirical knowledge and testing I’m not seeing this benefit and speed.
But for some reason people are claiming that. And it feels like I’m left behind.
Even if the real number is a 10% speed boost, I want to know where and how to apply this to benefit from this tools. It’s the same as the LSP, people can complain about it, and be like DHH that says that LSP aren’t worth it. But on my own usage it’s a good tool to add on my set of tooling.
But for some reason people are claiming that. And it feels like I’m left behind.
Classic FOMO, and that’s a large part of how they’re selling this snake oil.
I can feel the FOMO, but what makes me afraid is: “What if this isn’t FOMO?”
But to be honest, it’s hard to see the real benefit in the tool when you look on all the “broken” parts along the way.
The speed boost is only seen when AI is used by people who are already very skilled at whatever they are trying to use AI to assist them with. Code assist/generation LLMs only help in a meaningful, effective and efficient way when the user is proficient enough at coding to write specifically worded prompts and knowledgeable in the coding language to properly audit the output.
The trick to getting a speed and efficiency boost from AI is to already be good at what you’re asking it to do.
I’m not an expert in Game Dev, but I can see “good” code and “bad” code.
And yet, I can’t see the meaningful benefit, I can see the cracks, it’s never “fully” there, it’s always a couple details that aren’t correct in the planing, it’s always something that bothers, or some misleading thing that looks “ok”, but when you think little bit more it’s not trully correct.
I’ve spent 4 months where all my code was written by AI. I built my own harness, introduced a lot of deterministic checkpoint/guardrails, to try to ensure a consistent output from agents.
It definitely improved the output, but… It is still not all too great, if you want code that live up to your own standards.
It have taken a lot of my time to try to build my own harness, tailored to my way of working, but it’s difficult and time consuming. But I think that is what is needed to achieve a sort of reliable solution. The biggest issue is probably the cost of tokens.
A really small feature easily costs me $10 in tokens, if I am doing it the agentic way, and that is including smart model routing. These 10 dollars might have saved me a couple of hours, where I can work on something different.
Sometimes it is worth it, other times it is not, and it kinda sucks to be reliant on others harness’s like Github Copilot, OpenCode, Claude Code, because each update changes something that might interfere with your own setup.
Oh. I didn’t say it is useful for full human code writing replacement. Just that ANY speed boost would only be achievable by someone who already has the skills required to do the job quickly and accurately.
