Matt

Finding that hooks are so much more efficient at providing context dynamically than skills

@cursor_ai coding with composer is uncomparable to any other model, super excited to give it another shot

Composer 2 is now available in Cursor.

@nicoalbanese10 Building itself in the sense of creating features autonomously or following a PRD-like doc ?

I'm working on something that is now building itself and it's a pretty wild feeling.

agent teams are cool

@jarredsumner wow very curious about this since setting up windows tasks is still a mess

In the next version of Bun `Bun.cron(file, schedule, name)` schedules a function to be called on a recurring interval

20B$ AI can one shot me full architectural restructuring but will collapse on the simplest scrolling bar layout shift that i will ask it to fix

Convex is the only way to build sanely

@thekitze congratz man

brrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrrr 🦞 u can do it too anon just put your head down and be useful for 10+ yrs so ppl trust you enoughŃ. open source, articles, conference talks, workshops, meetups, gazillion useful tweets, youtube videos, free apps, etc etc etc. fkin hell man

@saltyAom I love it great work from you/the team ! Is there a future where our built-in types auto-validate input at runtime ? That would be a craazy feature to have

This is the main point why Elysia exists Not performance, not fancy stuffs This problem is very real and I learnt it a hard way Type signature from your actual code IS the documentation not the other way around We spent ~3 years to perfect this

Just ecountered the trickiest "bug" on @tan_stack table + react compiler where console log were the reason why my table was working and removing them was the reason it would break down lol

@mattpocockuk how to handle errors gracefully would be nice, but maybe out of scope for typescript

Everyone's been asking me to make some kind of TypeScript skill. It won't be a "TypeScript Best Practices" skill, because I think skills are best used for a purpose, like: - Migrating from JS to TS - Fixing spaghetti types - Improving TS perf - Extracting types from libraries What else would you want from this skill/set of skills?

@saltyAom you're great, my girlfriend loves when I get to use elysia because she finds it very cute (though it's a close contest with Deno according to her)

Thanks for staying with me despite how weird I am Thank you guys

@robj3d3 crazy numbers well done

I tested every affiliate strategy that actually worked for 7-figure brands. Then turned "just another referral program" into a growth engine that added $10k MRR in 30 days. 5 tactics you can copy. Zero theory. Use them to scale 👇

@gao_jude Not the same at all imo

@MattM_lj Yeah almost like pairing the main agent with a subagent that takes all previous context + docs and generates a list of relevant docs.

@gao_jude hey really nice article about skills, super interesting did you run an eval that only strongly tell the agent that the docs are all available in .next-docs/ without providing the index? would be curious to see the result

commands > skills

@gao_jude to me skills don't work because you distract the model that is supposed to do the work with other actions

@gao_jude I feel like we need to decouple the "what to load in context" with the model that does the job itself True it may get expensive to actually know what to load, but imo a model like composer can do this in 5 seconds quite accurately, and then you feed to opus the right context

@theo skills fall apart at scale imo I find myself invoking them on a prompt basis, future is probably an pre-agent advocating which skills to invoke based on context itself

You don’t need a bunch of MCPs and skills. You don’t need a customized TUI with tons of plugins. You don’t need custom sub agents and Ralph loops. You just need to prompt better.

@gao_jude thanks chef I guess listing docs is not enough to know what's in it worth considering an approach where a lighter model assess which docs to check for the task at hand ? and indicate the real model to check them would maybe avoid bloating agents.md at scale ?

@MattM_lj I got only 80% pass rate with this approach, specifically failing agent-029-use-cache-directive, agent-032-use-cache-directive, agent-037-updatetag-cache, and agent-038-refresh-settings.