I've been using Opus 5 for coding a lot lately, and compared to Codex (which I never had trouble understanding), this issue is driving me up the wall.
So much so that every time Opus 5 finishes a plan and shows me the summary, I have to prompt it again to explain everything it did in an “ELI5” way so I can understand what the heck it is saying.
For example, Opus 5 told me in a summary that “Net legs ran small rosters, not the plan's full ones”. When I asked for the ELI5 of what that meant, it said the client dropped video frames during testing and got 30 FPS instead of the required minimum of 60 FPS.
Sadly, I don't think so. I remember Leo saying some not-so-nice things about John on the pre-show to TWIT a few months ago. I never understood how Leo couldn't forgive John for whatever he did, but Leo has had long-lasting feuds with a lot of people over the years. All that wasted time fighting between those two when we could have had them doing great shows together over the past few years and giving us great content. What a waste of time fighting.
Too bad him and Leo had a falling out years ago. I really loved watching those two together. I remember John was always borrowing Leo's equipment, such as camera lenses, and he would always forget to bring them back and Leo would always scold him for that.
Not a blog, but a podcast, but I love "The AI Daily Brief," and I listen to it daily because it keeps me up to date on almost everything happening in the AI universe.
I'm sure this is not new thinking, but it was part of my discovery process. My objective wasn’t to crank out an app, I wanted to understand the process of using an agent to build something. I went into it thinking about how I’d work with a new employee. When I work with someone new, I explain ‘my API’ and try to understand theirs. How do I communicate, what’s my level of understanding, how can we collaborate most effectively. That’s been refined into the working_with_me.md doc.
I knew the context.md file, but read how big and cumbersome that file can get, and that as memory fills in a session, more of that context can be left behind. My project wasn’t about speed; I knew I would need a way to pass the latest session details - status, open items, decisions made in session, etc - to the next agent in a new clear session. That was the thinking behind a session state doc.
I created those two docs outside of claude.md, with session.md living in each project directory, and working_with_me.md available in the claude root directory.
My starting instructions are for the agent to read those docs - plus other project info, and my ending instructions are to update working_with_me and handoff.md. If mid-sessions something is lost - like the agent gets sycophantic, I’ll stop and dig into that, trying to understand how context was lost and how to firm up the point.
Very impressive, especially since you did it solo. The website looks great and explains everything in detail.
Can you elaborate more about its development? How much do 110B tokens equate to in $$$? What LLM did you prefer most during development? Any suggestions for other solo developers trying to launch their LLM-built product?
How are you accessing their API? Through OpenRouter, or direct? Are you using DeepSeek v4 Pro? $2 seems a lot cheaper than my own experience accessing them through OpenRouter for over 100 million tokens, but I am using OpenRouter to access v4 pro.
Cache hit rate dominates your total cost calculation for long agent session, and it largely depends on the provider. Deepseek's native deployment is probably much better than third party in this regard. For v4 pro it's a whopping >100x price difference between normal input vs. cached input tokens.
I am using Flash and accessing the API directly via vscode insiders and occasionally Zed (it's buggy but I keep coming back to it because I want it to succeed).
Unless you need enterprise multi-model management, I don't see the point in OpenRouter as it just adds cost overhead and you can just self-host an open router alternative (LiteLLM, Bifrost, etc). Running an LLM gateway locally is kind of nice as it allows you to normalize your configurations against your internal gateway - but I haven't really needed to.
Deepseek direct is atleast 2x cheaper than other providers serving it. Maybe their caching strategy is just significantly better or they’re subsidizing api pricing. I think it’s the former. $2 is closer to 50 million v4pro tokens in my experience
Ensure you're hitting the DeepSeek provider via OpenRouter - they have the massive cached tokens discount. If you're hitting any other provider you're paying an order of magnitude more.
Pro is substantially more expensive than flash. In addition, there's wide variance in price with DeepSeek themselves providing the cheapest tokens last I checked (but they train on them). Caching policy also varies by provider. TTL can be as low as 5 minutes or as high as 24 hours and reading from the cache might or might not reset the timer. Whether or not you get a hit makes (IIRC) a 10x (edit: it's actually 50x) price difference in the case of DeepSeek themselves.
This is so cool. Awesome job! I grew up playing submarine games: Gato, Silent Service, The Hunt for Red October, Up Periscope!, and Aces of the Deep. I wishlisted it on Steam and will be purchasing it when it is released.
Can you talk about the development process? How long did it take you? Is any of it vibe-coded? As you already stated, the browser version is a Vite/TypeScript app, but what programming language is the actual game in?
It's taken about 3 months of my spare time (some evenings and a few weekend afternoons) so far.
> Is any of it vibe-coded?
I actually started the project back in February as an experiment to train myself in prompting and to see how far I could get using Codex and letting it do literally all the coding for me. It turned out, I could go all the way. I have not touched a single line of code myself (I have only tweaked some CSS from time to time).
This project has led me to realizing my 20+ years of professional experience in coding are clearly worth far less than they were just a couple years ago...
> As you already stated, the browser version is a Vite/TypeScript app, but what programming language is the actual game in?
The actual game is written in TypeScript. For the Windows/Mac/Linux/Steam Deck/iOS/iPadOS/Android builds I'm using Electron to package it (which has surprised me for how well it's working even though I always kind of hated Electron...)
I would recommend WindowBlinds to achieve the "grey, boxy UI" look. As for a Windows 2000 theme out of the box, I am not sure, but I know it can make your Windows 10/11 UI look and behave like Windows XP.
There is a custom skin editor as well, so you can tailor the look of Windows to anything you choose, so you can probably get very close to the Windows 2000 look you are seeking.
Cheers. Some of the themes look pretty good! I used to use StarDock Start and StartIsBack back when I was using Windows 8 to, well, get the Windows 7 Aero theme and the start menu back.
That being said I do notice that many of the rounded corners aren't fully transparent...
So much so that every time Opus 5 finishes a plan and shows me the summary, I have to prompt it again to explain everything it did in an “ELI5” way so I can understand what the heck it is saying.
For example, Opus 5 told me in a summary that “Net legs ran small rosters, not the plan's full ones”. When I asked for the ELI5 of what that meant, it said the client dropped video frames during testing and got 30 FPS instead of the required minimum of 60 FPS.