Cristian

I don't understand why people keep mentioning "skills". Almost nothing is just a skill. A skill is just a saved prompt, that's it, so you don't type things twice. A SEO system like this involves many scripts, workflows, APIs and more linked together. The "skill" is just a way to call/trigger those systems, and define what the agent's goal is.

Woah, cool, I assume that's Opus? This is the best I could get from Astra, after a lot of prompting

I feel like something having a too detailed plan hinders implementation, because some issues are noticed only when implementing, and that drifts away from the original plan. I like to keep the plan a bit more high level, run the implementation step, and then test the app myself and glance over the code to see if I see any obvious flaws. Before I cared more about getting the code right from the first try, because implementing was costly and slow, but now LLMs can easily refactor or change architectures, so I prefer to start somewhere, and then tweak that architecture towards my desired optimal solution, instead of going directly for the optimal implementation.

Ok, I'm ready. Where reset? :(

Muse Spark 1.2 seems to be an incremental improvement. Slightly smarter, slightly cheaper and considerably (30%) faster.

I often choose GPT/Gemini/Claude thinking that the model is more stable, as it comes from a single provider, as opposed to using OpenRouter open-weights models ran by various providers, which might serve quantized or nerfed variants, but now it looks like we are just paying for black boxes and we have no idea if we are getting what we are paying for. There's also not really any way to test it, each request could be different, than can detect benchmarking code, etc. Models should have a "digital signature", a way to prove that this response was indeed generated by THAT model, at THAT reasoning level.