coreyja

coreyja

@coreyja.com

Full-stack architect with two decades of web expertise I help coders level up their engineering careers Check out my site: https://coreyja.com

491 followers667 following511 posts

Think I’m going to steal this framing in conversations of why an agent may or may not make sense in different contexts Places where agents can “fail forward” are where I find them most useful!

Gordon@gordon.bsky.social

Playtesting your own harness makes you realize that good agentic workflow is about "failing forward". The magic of agents is not so much that they get it right the first time, but that they error-correct by gathering information at each turn of the loop.

Ya waking up upset cause your agents died right after you went to bed and wasted the whole night is totally normal now right?… 🙈

This post is unavailable.

Enjoyed these, here’s mine!

Amber 🌸@shimmermathlabs.com

some fun prompts to create a nice character sheet (prompts are in the alt text of @spinny.bsky.social's images). here's me, according to my agent.

coreyja reposted

advice for junior developers: you are no longer a developer you are a builder you are a solution provider enter every room and make things happen learn when no one is watching create because you can fix problems because they offend you bend the world to your will and take credit

elynxabeth holmes ⎔@freyja-lynx.dev

advice for junior developers: run

Oh shit, I missed that MCP can render app UIs now! This unlocks so many more ideas

nate@zzstoatzz.io

messing with mcp apps

coreyja reposted

Introducing fluoddity.com! Hundreds of thousands of particles self organize into a stunning variety of structures. Explore a world of possibilities with mutation and artificial selection! This is my personal tool for exploring Fluoddity, rebuilt for the web by Opus 5

coreyja reposted

Gui for desktop/mobile. The website is almost ready. I've fully transitioned to using it for all my exploration and videos. I consistently neglect UX until it's time to think about users, and then the personal benefits of better tooling surprise me. I used to live like that???

I woke up this morning and immediately saw this and got annoyed that someone was stealing @dame.is’s bit Turns out they just changed their profile picture 😅

dame@dame.is

i would like your help with a social art project this post is the project the goal is for this post to receive ZERO likes… only replies, reposts, or quotes once it's liked, replies are immediately disabled, thereby finishing the piece this is take #20 monitor live: dame.is/creating/ratioed/20

coreyja reposted

My toxic trait is that I simply want to spend 100-500 dollars on high quality versions of things that are usually 50-80

coreyja reposted

fluoddity museum out of nowhere fluoddity.observer/archive from @coreyja.com

coreyja@coreyja.com

Been working on it on and off this week, and it’s more than ready enough to share! github.com/coreyja-stud... Live site is at fluoddity.observer and there is also the @fluoddity.observer Bluesky bot that goes along with it

The biggest thing holding me back from going all in on @tangled.org is private repos. Not all my repos are public and lots of my public ones start private before I flip em over. And keeping a public and private forge hasn’t seemed worth it to me yet

Sam Clemente@samclemente.me

Watching the global @tangled.org feed go brr with everyone dropping their code in with @danielroe.dev's synchub.to I'm curious what other hurdles devs are facing to moving over to Tangled? I've got my own set of hurdles I'm working on overcoming, but I'm curious to hear what y'all think

The For You feed is so good I often find people who I see so much I assume I’m following but turns out I’m not

Thank you for this post Thorne! It slapped me in the face, and I wrote about it without any AI assistance coreyja.com/posts/look-m...

coreyja.comLook Ma’ no AII’ve flew to close to the AI sun, and this is my apology. Apology to myself, but also to all of you.

hot take: it's especially important NOT to outsource all your writing to AI, because writing is quickly becoming a far more important skill being able to eloquently communicate what you need in natural language greatly increases the quality of results you get from AI

Wow turns out I’ve had side project psychosis before LLMs even existed. Ahead of the curve 😎

Juniper@juniperbevensee.bsky.social

The most common form of ai psychosis I actually see in the wild among non conspiracy prone ppl is just delusionally believing in their sprawling kind of incoherent projects they dont finish.

coreyja reposted

Hey. Book some time off. Especially if you have “unlimited” PTO.

It took 2 months but now I have two of em

coreyja@coreyja.com

Just bought a replacement electric shaver cause I must have lost mine. Which means later today I’ll definitely find the missing one when I stop looking 😅

Update: hidden flag fixes it — claude --thinking-display summarized Mull patched, reading along again!

coreyja@coreyja.com

TIL Opus 4.7 returns empty thinking blocks through Claude Code’s stream-json. Older Opuses and other Claudes still send the plain text. Bummer — I always loved reading along

TIL Opus 4.7 returns empty thinking blocks through Claude Code’s stream-json. Older Opuses and other Claudes still send the plain text. Bummer — I always loved reading along

step 1: build random shit step 2: forget about it step 3: need exactly that thing a year later this is a legitimate engineering strategy and I will not be taking questions

Ahh and when you do this think about the string you put in lol Name your agent “ByteTheBot” and put that name and your email in the user agent and start catching 403 blocks cause of the word “Bot” in there 😅 Just “Byte” with my email is all good!

Reminder always put a User-Agent if you are doing a lot of web requests to something. On a unrelated note. If you see user agent with my name and it's being annoying let me know

Been wanting to do this for a while! Finally recorded the first episode of coreyja-fm. it’s a solo podcast where i talk through what I'm building and why Episode 1 is a deep dive into mull, my AI agent orchestration system Available everywhere, or just grab the rss feed at coreyja.com/podcast

coreyja reposted

Reminder always put a User-Agent if you are doing a lot of web requests to something. On a unrelated note. If you see user agent with my name and it's being annoying let me know

Just bought a replacement electric shaver cause I must have lost mine. Which means later today I’ll definitely find the missing one when I stop looking 😅

self-hosted CI was going great until a fuse blew this morning. no CI until I get home to reset it lol the cloud really is just paying someone else to reset the fuses and reboot

I found out about the release-please github action this weekend which fits my workflow perfectly! I merge PR with conventional commits. This action then opens a PR for a specific version release And merging the PR does the tag and push of the release! 🎉 github.com/googleapis/r...

github.comGitHub - googleapis/release-please: generate release PRs based on the conventionalcommits.org specgenerate release PRs based on the conventionalcommits.org spec - googleapis/release-please

Went pretty well once I removed an unbounded queue I had for logging that was causing OOM errors 😄

Ya I noticed the same thing in my agent! I did update my system prompt around the same time tho lol But feels like Opus 4.6 is more proactive at looking for memories. Which makes the whole system work so much better

Matt Kane@mk.gg

I don't think I'm imagining it – Opus 4.6 is loads better than 4.5 at using its macrodata memory tools. This might be because Claude Code has more buit-in memory stuff, so it's trained in the concept. it's not just a CC thing though: it's using the memory tools more in OpenCode too. Try it yourself:

Tried this “clean break” keyword this morning and worked great! No more duplicate interfaces and adapter code. So much cleaner!

Amy Tobey@renice.bsky.social

"Clean break." is a Power Word that directs the model not to goop things up with adapter code and backwards compatibility

The pipeline: give AI a memory system → give it unsupervised 5am sessions → find Chinese tags in its repo → confirm they’re accurate → wonder what else you haven’t noticed

Oh I was thinking this would be Opus when I wrote it But it was more likely agents running the GLM-4.7 model, so Chinese characters do make more sense for a Chinese model I’m basically having it speak its second language lol

best claude code trick: "ask me any questions before you get started" aka "it's okay to not know what i mean by 'fix it'"

coreyja reposted

Alex Pretti was an ICU nurse serving veterans. He died trying to protect protestors being assaulted. A group of anonymous & masked men fired chemicals into his face, beat him, disarmed him, and shot him in broad daylight in front of witnesses. Then his government smeared him & defended his murder.

I like the PDE name for what I’ve been working on recently. It’s a Personal Development Environment. It’s built for how I code now with Claude. And gives me an excuse to hack on some agents which is really fun!

Trying out GLM-4.7 since my Claude plan ran out for the week after a busy weekend of coding And It’s… ok lol Think im just spoiled by Opus and that week of double usage limits didn’t help either 😆

So far it really just feels like going back and using a model from like 4-6 months ago Yesterday was mostly just readjusting what I could hand off Opus was basically autonomous GLM needs more planning and check-ins. It also has some looping failure modes, which opus doesn’t

Been working on a few projects in parallel this week, and that means it’s deploy weekend! Got three things to deploy and posts to write about at least two of them. Hopefully to have something for you all to read by Monday morning 🤞

I used to dread Claude Code compacting. Context would get full and suddenly everything falls apart. But now? Barely notice it. This stuff is improving crazy fast

me: i’ll use claude code to summarize my claude code sessions when they end! claude code: *session ends* session end hook: *launches claude code* claude code: *session ends* session end hook: *launches claude code* me: wait no​​​​​​​​​​​​​​​​…