@KSimback

Can token spend awareness actually cause users to be more efficient with token spend? That's what I'm testing with a new lightweight harness called Gauge When you prompt Claude Code, it runs a quick prediction on expected token spend If it detects your prompt is unusually inefficient, it gives you a warning and a chance to edit your prompt with some helpful tips Otherwise it runs normally but still gives you an expected token spend + final result of actual vs predicted It also sneaks in a "pro tip" to help users become more efficient Dropping some examples in the replies

Putting an LLM as judge is the simplest and easiest way to create a loop 1. Do the thing you already do by prompting AI 2. Use another model to evaluate the output and give it feedback 3. Iterate until both are satisfied Only see the final output once the loop is complete Since you're using an LLM as judge, this is generally best for non-deterministic tasks like research, design, building specs, etc. Using Looper to do this will make it super easy by refining your prompt for a better constructed loop, adding the judge, and setting the exit criteria so it doesn't run unnecessary iterations Now at 500+ ⭐️ https://t.co/smAXbBsudu

We’re at a point in the AI cycle where first principles thinking will be rewarded

The next 6 months are going to be entertaining > Grok 4.5 + the 2T run > Fable / GPT-5.6 release drama > Anthropic IPO > GLM catching Fable by Q1 > DeepSeek pushing new innovations > Nvidia’s local AI push Open question - where is Gemini and Meta in all this?

I love the recent debates on memory, but here’s what many are missing: In the span of ~6 months we’ve gone from mostly general chat to heavy agentic usage that is 5-100x+more memory intensive And it won’t stop there The demand side of this equation is insane https://t.co/MFm4ohjLQL