Yogev (@YogReport) header
Yogev (@YogReport) profile picture

Dad, husband, tired. Writing music for video games (Mission Zigloton, HipWitch, The Siege of Jeomdo). Building some cool stuff using AI.

London
Joined May 2009
Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)Photo by Yogev (@YogReport)
Two days ago I wrote "Hey, Gemini Is Usable Again!" after months of not using Gemini. I was happy with Argon but its quota was restrictive. Gemini 3.8 Flash was the real surprise: good for smaller coding jobs and a much more generous allowance.
1
33
Previously I had a capable model I had to ration heavily, and Flash for simpler tasks. Now there's a capable model in the more generous pool. There's also Orchestrator: Neon plans and reviews, Hydrogen implements. Haven't tried it yet.
1
12
I get Google AI Pro through my existing Google Drive subscription, so no extra cost for me. At the end of my last post I hoped Google would increase Argon's quota. Instead it looks like they've given me another model to use in the more generous pool.
17
When I started using AI, I tried the free versions of all the providers. Gemini came with my Google Drive plan, so I used it alongside a paid basic Codex plan. Codex mostly for planning, Gemini for building. Around £20/month.
3
2
77
Flash is amazing for simple jobs and scoped coding tasks. And with its usage quota, it feels like it can go on forever. Gemini 4 Carbon is out too, but it's still Ultra-only for me. Hopefully I can try it on my plan eventually.
1
61
So I'm back to using GPT, Claude and Gemini together, switching between them again depending on the job and how much usage I have left. I'm getting quite a lot done with that setup.
33
My guitar practice problem wasn't that I needed more learning material. I wanted one place that remembered what I was practising, how long I'd spent on it and what happened last time. Oh yeah - and streaks. Streaks are cool. For an invite to the alpha - drop me a message.
1
35
I asked Codex to investigate why my home network felt so unreliable. It looked at the router's actual admin page, topology and logs. First discovery: a mesh node upstairs had been switched off. Whoops. Well, that’s another thing AI can do for you.
1
52
One thing I like about building software with AI is that I can make things just because I want them. I don't need to convince myself that an idea could be a £100m company before I spend time on it. Sometimes 'I wish this existed, exactly like this' is enough to start.
19
At the beginning of the year I presented my (then) MyClaw team at work, including the custom HQ I'd built for myself. At first people didn't really understand why that would be useful, but after seeing the real work examples they were super excited.
1
27
I've done the same with my own projects (where I'm the one paying for the tokens). QuotaBar tracks my Codex/Claude usage the way I want, and I built Rephrase because I got tired of copying text into ChatGPT and back just to improve my phrasing.
1
15
For coding I built Baton so Codex and Claude follow my build/review process - one builds, the other reviews, and Baton handles the handoffs. I don't need to adjust the way I work to my tools anymore. I can adjust my tools to work the way I like them to :)
11
More and more coding interviews seem to be moving away from LeetCode (yay) and toward “here’s a codebase you’ve never seen—find the problem and fix it.” That’s actually quite hard to practise alone. You need an unfamiliar codebase with problems you don’t already know about.
1
20
That’s another great use for AI: ask it to create a codebase with problems for you to fix, without telling you what to look for. Then have it review your work and provide feedback.
10