It’s not just a prompt. The agent SDK actually has real mechanics for it. You add the metadata, and the system only pulls the skill in when a task matches. Then it follows the instructions you’ve written, including any scripts or files you’ve attached. That’s the part that makes it different.
It’s situational and reusable and only known when needed, not perpetual.
I think most of that comes from how they operate, 4.8 was straight to the point operating from in chat context and memory making it more fluid, then 5 does more self-verification, re-reading text rather than operating only from memory and context, making it seem as if it has no recollection, losing the fluent continuity and making it much slower, it’s not worse in properly guided agentic workflows though.
Python is used for orchestration and I/O, the heavy compute runs in compiled C++/CUDA binaries. Doing all of that high-level work in pure C++ would be painfully slow for research and development. There’s a point of diminishing return, you wouldn’t notice a difference in speed and development time would be crippling slow with no benefit.
No I’ve seen how people code with AI and it’s completely wrong, while it may be effective in some cases, it’s inconsistent.
You have to codify the way they behave, create templates, implement automated gates based on codebase scans that detect format drift and blacklisted patterns, security compliance scans, etc.. those blockers then provide feedback so it self corrects.
Combine that with skills and main instructions on how to interact with the repository, it works quite effectively, but as always invest in yourself, while it’s working for you get more familiar with architecture.
With some practice and review, absolutely. I admit, I have gotten a bit lazy with AI but I still control decisions, design and logic.
Fun thing to do, create a reusable functions or modules and unit tests, have ai do the same thing. See what it does different, whether better or worse you learn from it.
@logan14510@HalimehN13189@Bober_smart Yeah I don’t see this doing well unless it’s properly segmented, and categorized per vault. 100% on RAG, but only if implemented correctly, it’s also more expensive accounting for embedding models and techniques, and reranking, enrichment.
@Tectone Bahahaha, imagine Chipotle family dinner while playing Tectone stream in the back as he reacts to the lover birds. Jokes aside, this looks so fitting, like it’s meant to be, truly a match made in heaven.
@ni_dendo Aside from the slim chance of this scenario, this is not how this would turn out at all, that dog would have gotten dragged up a tree somewhere lol
@ascensionG2026@Garnet_2203 I would think that being inside your own home is more of an exposure than an EV, actually sitting in a car while the sun is out is much worse.
@DimpleAroro23s My windows machine is a fully debloated offline steam machine, with just enough to connect to steam via proxy and run Sunshine for streaming. Otherwise been on Linux for a very long time.
@Hello_Ashleyyyy Pillow fighting, just found out it’s an actual sport after I noticed the TV at the gym showing two grown men in a boxing ring spin slapping each other with pillows