As I understand it, cutting through the marketing blabla, It's for an agent (as an LLM-in-a-loop) to iteratively improve its own prompt based on initial instructions, thereby saving you tokens and time when writing prompts.
Ofc the LLM can still ignore relevant ones, load the wrong one etc
They do count when they are used, but the idea is that they are used only when relevant so if you compare this to sending all the instructions all the time you'll end up saving tokens.
Well I think you could add shell scripts and the like, it's just very rarely done apparently. Also, I would like do make it very I clear I do NOT recommend downloading and running random shell scripts on one's computer if one does not understand what they are doing.
Yes they are basically parts of prompts that can be added ad hoc, thereby in theory saving you tokens compared to just sending a full prompt every time.
Yes, hello, is that God? Yeah, hi. I have a complaint. You gave me this brain that’s designed for finding berries and avoiding lions and now people are ‘just circling back’ to see if we can ‘move the needle’ on ‘key initiatives’? NONE of those things are berries.
Evergreen
I get a lot of mileage out of that quote for sure!
Portrait of Upton Sinclair
Personally surprised I am having a hard time making someone understand this who's salary depends on them not understanding
That makes sense to me and is also roughly where I landed when experimenting — manual intervention ofc seems hard to square with the hype b/c it means you still need a lot of skilled software developers to get sth done in the end
If you say refactoring, do you type those changes by hand or give instructions to the LLM?
Fucked up a perfectly good transformer-based translation engine, look at him, he's got anxiety
"It’s been a while, so quote this skeet with your favourite $20 dollar word."
autochtone
It's more efficient b/c you don't actually have to throw it away! Synergy!
Important to note that the difference between a design prototype crafted by hand with HTML, CSS (what I aim for and like I believe @set.studio offers) and a vibe coded prototype is that you can actually use the code from the first one for the final product.
I am currently writing down my thoughts on coding with agents, and yeah that feels like a great one-sentence summary of the first 1000 words
So you save time with relatively terse instruction but the parrot still behaves
Yes after dabbling I am pretty confident most of the reason why LLMs seem better at code is clever product engineering around the model.
All the "context" (tool calls, mcp, skills) constructs a prompt for you that's much more likely to get you the output you want
From a discussion in RISF based on the old IBM adage, an updated version for the modern era:
Nice
I feel the same way, I sometimes see this floating around as advice for "semantic" styling, but in effect you're tightly coupling HTML semantics and visual meaning, with the effect that every project of even moderate size has to overwrite 80% of its heading styles.
We should do a contest with a panel of judges scoring who can get the most sycophancy out of it
Well sometimes I use this to be polite when I really want to say "I don't want to build this because it is bad design qua design and will make the product worse" (such is my arrogance) but point taken that I should find a better way to say this
But also Apple's fault for hiring someone with the completely wrong profile
Oh definitely, I missed that post the first time around!
The fun bit about this is that he was allegedly poached, not fired
🔥
"So, if 'pixel perfection' is no longer a viable goal, what should we be striving for? The answer, I believe, lies in shifting our focus from individual pixels to design intent."
Some more 🔥
"A pure 'pixel-perfect' approach treats every instance as a unique snowflake, which is the antithesis of a scalable, component-based architecture."
This is fantastic by @amitsheen.bsky.social: "True perfection means a site that works for everyone. [...] 'Pixel perfect' often prioritizes visual aesthetics over functional accessibility [...]" 🔥
www.smashingmagazine.com/2026/01/reth...
Orchestration as in for coding agents? I would say it's the exact opposite, breaking down tasks needs a lot of hands-on technical knowledge and management without that is generally horrible at it
LLMs in this case are a shortterm and likely unsustainable bandaid for issues that are already caused by underinvestment and shorttermism, mostly appearing in proximity to bullshit tasks (following Graeber), i.e. where we do box-ticking, duct-taping or taskmaster-responding
