“The most important skill in prompting is expertise in the domain you’re prompting for.”
LLMs reward expertise
seangoedecke.com
Tom Darby
@tom.hitlst.io
Building HITLst: process on rails for human-agent dev teams. The thing where agents do the work and you still ship something good. → hitlst.org
“The most important skill in prompting is expertise in the domain you’re prompting for.”
LLMs reward expertise
seangoedecke.com
Film you've watched more than 6 times, with a gif. Hard Mode: No Star Trek, Star Wars, LOTR, Marvel, Disney, or Ghibli.
Zorg Says My Favorite Meme
ALT: Zorg Says My Favorite Meme
static.klipy.com
Film you've watched more than 6 times, with a gif. Hard Mode: No Star Trek, Star Wars, LOTR, Marvel, Disney, or Ghibli.
Post a perfect album from the 90s that isn’t Nirvana, Soundgarden, Pearl Jam, or Alice In Chains
Post a perfect album from the 90s that isn't nirvana, soundgarden, pearl jam, or alice in chains.
Yes, I would like a refund on this air you sent me today No, it’s the wrong color, and the smell is off Thank you
Wait, you… use libraries? Code that other people wrote for you? You just… download it? Do you even read the code? Do you not care about the craft, the knowledge that comes from putting in the work yourself? Aren’t you afraid your skills will atrophy?
It’s not the typing func doThing(blah blah) that does it for me, its the knowing that the interlocking components work in just the right way to make everything else hum along nicely. In other words, I do understand the magic of the machine. No coding required anymore.
I spent decades of my life loving the craft of writing the best code, priding myself in generally being regarded as pretty good at what I do. LLMs gave me a bit of a forced perspective shift that taught me what I really enjoy is the building of the system, the architecting of how it’s going to work.
Don’t you *LIKE* understanding the magic of the machine? You do realize hand-programming (I hate that I even have to specify hand now) is fun…right?
Over the years: Kurosawa, Gilliam, Jeunet, Besson, Wood (for a Windows VM used solely to compile things on Windows)…and a bunch more I’ve forgotten, but you get the drift 😅
Currently have the chance to name things. How do you name things. How do you dodge both cliche and jargon. How do you hit the right level of wonka not too cute but not too dry. Gotta have the right name. Gotta feel good on the tongue. Evocative but not horny. Maybe a little horny?
My spouse has been living with breast cancer since 2012. It’s been metastatic for over 5 years. I…very much disagree with this sentiment, in a profoundly visceral fashion.
Emma Pierson could directly benefit if AI cured cancer. Yet, she writes, as a professor of AI who was once mentored by Anthropic’s co-founder, she’s rooting for AI progress to slow down.
As someone who is NOT a coder, I've spent the last two weeks building some things from scratch using Claude and Python. I know this is a somewhat controversial take, but for people with my skill set, it rounds me out somewhat and it has been amazing.
Didn’t get much time to work on HITLst this weekend, but Fable was able to put enough well-scoped work together before they got zapped that I could let Opus chug through the backlog relatively unsupervised
Nextdoor is something else though
Made the mistake of logging into Nextdoor and discovered some amazing things, including the person in my neighborhood who thinks Tesla is going to make Uber, Lyft & Waymo extinct... and someone else who thinks we need to outlaw swearing in public.
The Fable/Mythos news tonight is…not good. If this is the response now, what’ll it be in a few years when you can run an open model as powerful as Fable on your own hardware?
So much of effective use of AI coding tools is managing your own cognitive load. Learning where to focus your attention and where to trust the machine to know what it’s doing is probably the biggest factor in successfully using these tools. If you don’t, the volume of code will overwhelm you.
Ultracode + Fable:
Man with Flamethrower: Kill It With Fire!
Alt: Man with Flamethrower (from “The Simpsons”)
static.klipy.com
Ultracode with Fable seems to be risky, so I'd caution against it. What I've seen is, Fable meticulously plans its work in batches (good) but then Ultracode runs its workflows (good) on each stage (bad). It's overworking the task and will burn serious cash doing it
alright, Fable is sick. It *just* gets what i mean went through some kind of complex design, off-the-beaten-path, and it one shot implemented in 40 min with a workflow to check its work and it did the right amount. Not too much, not too little. And it works, very well
OK, second evening of working with Fable 5, and I'm even more impressed. Given very general guidance–no more than a paragraph worth of text–Fable has been able to identify meaningful gaps in HITLst's UX, *and* propose and implement clear, easy-to-use solutions. Fable is a huge leap forward.
Oh, also: Fable 5 will absolutely *BURN* through tokens, but it's doing this as an orchestrator, not as the core model. Fable just launched five separate subagents to tackle five issues in parallel, which burned through roughly half of my Max 5x session quota in 15-20 minutes...
Gut reaction to my first evening's use of Fable 5: It's been doing some fairly deep dives for me tonight and has turned up things that Opus hadn't. I'm sure some of this is iterative improvements, but the overall experience feels both more effective and focused. Also, it sounds more professional.
I used it to make a homebridge plugin cuz there weren’t any great solutions to adding a state monitoring + on/off switch for my NAS to HomeKit. Total non-coder, was surprised the exercise actually yielded a positive result
This is the big problem I’m working to address with HITLst.io. Build quality enforcement into your everyday work, automatically. Learn as you go with your agents. Run ceremonies to make sure you’re not drifting. Keep the agents honest, do the work right.
HITLst — project management for human-agent teams
MCP-native project management for human-agent teams. Tracks intents, enforces gates, runs ceremonies — so AI coding agents ship with the rigor of a well-run engineering org.
hitlst.io
AI creates two quality challenges 1. How do you enforce quality output in models that hallucinate and fail to follow instructions? 2. How do you teach quality standards to people who lazily outsource their thinking to Claude and don’t catch the problems in #1 ? This is now where I spend my time
Yup. In 1 hour and 250k tokens, Claude cleaned up, organized, pruned, updated and finished implementing a Homebridge install I’d spent *weeks* trying to do myself. Everything works now, runs with the lid closed, auto-reboots and launches on power failure, no more terminal windows…everything. 😙🤌