Craig B.

@thekitze Whole heartedly agree and yet it’s such an obvious design choice. They could have done it the way you describe but then it’s wouldn’t feel as magical. Ppl pay for magic

skills infrastructure is incredibly dumb 1. there should be a boolean in the meta indicating whether we want this skill to always be loaded in context (yes, even the metadata of it). 2. COMMANDS != skills. i need ONE TIME taggable things that pull certain knowledge in the prompt, without preloading it or the agent guessiing it 1 conflicts with 2 in a way but currently its the worst of both worlds...

claude killed apostrophes

@tdinh_me fixed in new build, tnx @Stevie_658jjh... honestly this is pretty slick! well done mate

@tdinh_me best $9 ever spent, tells me i don't have node installed 💸... i use node every day, so, um... 😬 looks good though

This guy demoed his Claude Usage Analytics tool to me the other day at HRG coffee in Da Nang. Super impressed with the amount of data and analytics you can get out of claude code that I don't even know exists. (The color theme could use some improvements tho 😂)

There is a category of AI harness enthusiasts who will marvel at how good the review agent was to find the root cause of an issue that was caused by another agents code. And the original code will be something like a race condition that a human JavaScript developer would simply never make unless lobotomised

Tales from the delivery risk spectrum A) Every PR must be reviewed by agent B) Agent commits straight to git C) … and deploys to dev D) Agent edits the production code on the server A) <—- you are here

been using a monorepo since start of year because of this, and it turned out to be one of the best things ever

get your tap shoes ready !

@thdxr 10 years without a build status back to normal message is too long

who is gonna bring back stateful CI i'm ready to spin up jenkins again

seeing a github preview, interesting that agents are appearing under the user. maybe a lean into the metaphor that agents are team members that you lead

It’s fascinating the blurred lines between AI builders and traditional builders, those that were grinding long before AI. on the surface the outputs may be similar. but the later, has a real world system. and most importantly they exercise. the pervasive reality of grinding on a local AI harness , is that you can’t exactly disconnect if you are the operator . I would hazard a guess those who know how to ship, are doing everything they can to be able to orchestrate from their phone so they can still live in the real world, so they can still lift and run or embrace in the richer joys of life , all the things that provide the strength to be at their best when they are back to the grind. If you’ve grinding , but not creating systems and not disconnecting (this is most important ) beware !

If your AI workflow requires you to sit at the keyboard all day, you didn’t build a system. you built a grindstone.

Tragic temptation of the magic lever!

What is the best multi-player agentic development harness that you have seen?

@grok when i look at my profile in incognito mode, it says "@CraigBarrDev hasn’t posted" BUT this is a lie. Why? How to fix...

Here they are: Today: spa.ts0.org/2026/03/10/ Yesterday: spa.ts0.org/2026/03/09/

A guy I know with zero design sense is posting newsroom-quality interactive sites every day with a Paperclip company. That’s the shift: you no longer need personal mastery of every craft You just hire agents for the parts you’re bad at.

So slick, it feels like a gimmick Digging a bit deeper, it seems it’s got legs!

Early UI design work shouldn’t pollute git history. My Codex workflow: • Generate 4 variants • Use copy git apply • Test locally • git nuke • Repeat for each version Now you’ve explored multiple working designs. Only commit the winner.

English is the new code in practice