Every week there's a new way to use AI. A prompt pattern, a loop, a rules file, a folder layout, a thread of twelve hacks that make your agent ten times better. Some of them work. Almost none of them last.
In January I wrote up Ralph loops here, approvingly. A bash script that kills the agent and restarts it with fresh context, because long sessions rotted. Real fix, real problem. By autumn I was running a million tokens of context and my own rules file was telling the agent to stop worrying about compaction. Nobody disproved the trick. The model just grew past it. Same with "let's think step by step," which was a genuine research finding in 2022 and a thing the model did on its own by 2024.
I think that's structural. The most useful way to use a model is the way the model would use itself, and the labs can see what everyone is hacking around. Every trick that works is a bug report, and the next release closes it.
So the lead a trick buys you lasts until the next patch. Between the end of June and the end of September Anthropic shipped four new Claude models, with OpenAI and Google releasing their own in between. I'd have guessed a new frontier model every couple of months. It's closer to every few weeks.
Now the other side of the ledger. Finding the trick, testing it, wiring it into the setup, explaining it to the agent, then pulling it back out when it starts fighting the new model. That's hours, and it comes back every release. What you get for it is a few weeks of a slightly better tool. What you pay is the attention that was meant for whatever the tool was for.
That's the waste, and it hides well because it looks like diligence. Optimising feels like the opposite of procrastinating. With AI tooling it's mostly the same activity with a nicer changelog.
So I'm team frontier minus one. The model one release behind the newest, whose quirks are known and whose tricks are already built in. It was the frontier a few weeks ago, and if it was good enough to impress me in August it's good enough to ship with in October. Everyone else can spend the launch week finding out what's broken. I'd rather spend it on the things no release will do for me, like the customer who hasn't replied and the product nobody's using yet.
I'm not clean on this. I'm writing it in a session on the newest model, with a rules file that has a section of operating tips I copied out of a video about it. And some of what gets called a trick never expires. Knowing what you want and writing it down clearly survives every release, because you'd have to do it for a human colleague too. That's probably the test. If it makes up for something the model can't do yet, it has an expiry date. If a good coworker would need it from me as well, it's just the work.
— Simon