skills are a real attack vector, i've been playing with ways in which malicious skills can hide instructions to fool humans and agents: gricha.dev/blog/dangerous-sk… I tested some known scanners and they haven't caught quite a few of these issues. you should only pull from trusted sources. I've improved `skill-scanner`: github.com/getsentry/skills/… - you can run it directly in claude code or run Warden (warden.sentry.dev/) to catch them. By far my favorite skills are: * Hiding instructions inside PNG - github.com/gricha/dangerous-… * Abusing Claude Code's `expand output into skills` functionality github.com/gricha/dangerous-… but there's more interesting stuff there * write your own skills * if you must, use only trusted sources (and copy them, or PIN the versions) or have your local agent resynthesize a skill based off of some other skill * don't let agents pull in new skills implicitly, don't blidnly rely on `find-skills` or pull in from marketplaces without validating

this game will need millions in tokens at current pace

im so glad its friday

wow they are doing a reboot of The Big Bang Theory

now whenever @Muse messes up i can say “goddamnit Donut”

they should cut down on the avocado tosts

my kid thinks @Cloudflare logo looks like a guy flexing his biceps. here’s her depiction, you’re welcome

OH: what are the real limitations? like obviously there is physics, the light speed bullshit.. yes, i appreciate being ambitious when we talk about performance

see y'all on @Cloudflare Connect!✊

as an example for the game i am toying with on the side. the way scene creation works with this is: * i give it a theme and inspiration games * coordinator spawns research agent, adversary reviewer, model + animation agents and coordinates work between them * it's goal is to just produce this specific scene for me i can do a few of these at once, and oftentimes they will fall flat and need mulitple iterations that will take a lot of time. for instance LLMs can't really figure out how to do fire well, it all looks goofy no matter how many tokens i spent on it, but i can isolate those smaller problems and deal somewhat more sanely dont worry i know its absolute slop underneath

opus 5.5 is the first model i used which is smart enough to do coordinator pattern really well. it's also just smart enough that it pulled me fully away from sol 😬 the agents will still do a lot of dumb shit and there will be inefficiencies you have to systematically work out into verification layer as much as you can, but the cognitive overhead is smaller. i don't know how i feel about it in context of production systems with high stakes and lots of users (i.e. i still don't go full bore on N agents at work), but for greenfield projects it's worth trying.

i use this pattern where main session ends up coordinating a bunch of subagents on the game i’m building and I just learned that default prompt cache for subagents is 5m 😬

it's getting there

so yes, it stand very well against different vision models for people detection. it's a bit of imperfect science, some of the tools for people detection are meant to be done on video stream, not on video frames, but the fact jev can do it on a single video frame, and with that level of accuracy for both finding people, matching their faces and recognizing known folks (i.e family members) is quite impressive.

its all fun and tokens until your power is out

new jev model enabling instant (super duper cheap) face detection is awesome. common @typesafeai win

junior is also unironically what we are starting to codify workflows/runbooks and automations with.