Pinned Tweet
New Article!
Functional ≠ Declarative
I am sick of saying I like declarative code and being told to juST UsE ELm or HASkElL.
dev.to/playfulprogramming/fu…
Every week, my mind is opened to new ways I can use LLMs in.
I just added some audio from my phone to one of the videos I was editing and thought, “Hey, let's see if it can figure out how to process the audio to remove the background noise.” So I gave Codex the instruction to create a sub‑agent, and I had mentioned that my normal audio‑processing workflow used Rx Elements.
It freaking found Rx Elements on my computer, and even though they don't have a CLI, it found the installed components and found a way to use them directly anyway. It wrote Swift code and manually investigated the behavior of the components because they were undocumented. And then it built a CLI around them. It took less than 10 minutes.
🤯
I can finally automate the audio processing for my videos.
Funny how my agent team can make infuriating mistakes again and again and then suddenly get everything right all at once.
I wish I was smarter. There definitely is skill involved here. I hope I get really good at this. And I hope a new model doesn't invalidate what I've learned.
Is there better UX guidance concerning performance issues? This is what I have, somewhat inspired by Elon's algorithm:
I have so much to learn still about managing agents.
I have a "project manager" description that says to spin up subagents with specific roles... But they kind of weren't following their roles... Super frustrating.
To my horror, I just saw that I told the project manager to "Give [the subagent] the [Designer](DESIGNER.md) description."
🤦
I just changed that to "Tell it to strictly adhere to the [Designer](DESIGNER.md) description."
Hopefully this fixes the issues.
I used to think sprint retro was mostly a waste of time where people would nitpick. The major issues were already obvious during the sprint.
But I like having retro with my AI agents. I think it's probably critical, actually. They don't complain unless you ask them to.
Earlier today: "In both U06 and U09, Grok accepted the prompt but produced no file changes or reported error before the Team Lead stopped it and reassigned the work."
So I started an independent Grok session and... 23s?
I don't think Grok is doing okay.
Tip: Have Grok find all the UX opinions you've posted on X and then include those in your design prompts.
One of mine is that UIs that expand are a lot better than modals. I like both UI co-location and code co-location!
Wow, I was literally about to publish a video that says this exact same thing, even referencing the two pizza team rule.
You know what? I'll just upload this part here.
Not the final video, but I wanted you all to see what I was working on.
Tired: Two-pizza teams
Wired: Two-slice teams
@danshipper on the best team structure in the AI-era
Mike Pearson retweeted
software engineering is NO LONGER HELD back by coding, needing to acrually write things down, satsify complier, handle the details
as such coding IS dead, horray
software engineerring is unlocked to reach new heights – at Zurich, Switzerland
Guys, I am absolutely mind blown.
With AI demos, usually I suspect that there was a lot of back and forth prompting.
But the Opus 5.5 motion graphics demos, they're super real. I just made one myself, and this is incredible.
I hope this increases my content output dramatically
As with all AI generated stuff, I'm sure I'll see more problems the closer I look at it.
But I should be able to intersperse this within other videos, even if I have to manually design specific graphics that need to be high accuracy.
Here's the article this is based on. dev.to/playfulprogramming/fu…
Grok's writing style has begun to sound like Claude...
"It is not X. It is not Y. It is Z." But nobody was talking about X or Y.
Is it attributing its own reasoning process to me? That would make sense... All the LLMs seem to struggle with the concept of multiple people/agents. It's like when it updates documentation with details the readers do NOT need to know about.
It leaks the process into the final output. So it considered X and Y, and even though they are stupid and don't answer my question, it feels the need to talk about them because they're in context and it's not clear to itself whose idea it was, and what the user needs to see.
I wonder what the fix for this would be. Probably some reinforcement learning that sets up situations where various agents/people should not know details until they ask. Maybe this would reduce alignment. 🤷
This image is another example:
Great video from @fireship_dev
However, it's objectively more efficient to organize information declaratively. Frameworks are not going away. And TypeScript is still better than JavaScript.
Most of what makes DX better makes AX better too.
youtu.be/OuNKBjuV7A4
