Has anyone ever seen an instance in which the automated "How" removal actually improves an article title on HN rather than just making them wrong?
(There probably are some. Most likely I notice the bad ones more than the good ones. But it does seem like I notice a lot of bad ones, and never any good ones.)
[EDITED to add:] For context, the actual article title begins "Superpowers: How I'm using ..." and it has been auto-rewritten to "Superpowers: I'm using ...", which completely changes what "Superpowers" is understood as applying to. (The actual intention: superpowers for LLM coding agents. The meaning after the change: LLM coding agents as superpowers for humans.)
This is so interesting but it reads like satire. I'm sure folks who love persuading and teaching and marshalling groups are going to do very well in SWEng.
According to this, we'll all be reading the feelings journals of our LLM children and scolding them for cheating on our carefully crafted exams instead of, you know, making things. We'll read psychology books, apparently.
I like reading and tinkering directly. If this is real, the field is going to leave that behind.
I often feel these types of blogposts would be more helpful if they demonstrated someone using the tools to build something non-trivial.
Is Claude really "learning new skills" when you feed it a book, or does it present it like that because you're prompting encourages that sort of response-behavior. I feel like it has to demo Claude with the new skills and Claude without.
Maybe I'm a curmudgeon but most of these types of blogs feel like marketing pieces with the important bit is that so much is left unsaid and not shown, that it comes off like a kid trying to hype up their own work without the benefit of nuance or depth.
even if it works as described, I'm assuming it's extremely model dependent (eg book prerequisites), so you'd have to re-run this for every model you use, this is basically poor man's finetuning;
maybe explicit support from providers would make it feasible?
Curious what you think of sub agents, don't they still consume a massive amount of tokens compared to simply running in main context? I'm skeptical of any process that starts massively delegating to sub agents. I'm on Pro and don't think its worth upgrading to 200 a month just to not pollute main context.
I am not ashamed to admit this whole agentic coding movement has moved beyond me.
Not only do I have know everything about the code, data and domain, but now I need to understand this whole AI system which is a meta skill of its own.
I fear I may never be able catch up till someone comes along and simplifies it for pleb consumption.
I’ve personally decided that cursor agent mode is good enough. A single foreground instance of cursor doing its thing is plenty enough to babysit. Based upon that experience I am highly highly skeptical people are actually creating things of value with these multi-agent-running-in-the-background setups. Way to much babysitting and honestly writing docs and specs for them is more work than just writing parts of the code myself and letting the LLM do the tedious bits like finishing what I started.
No matter what you are told, there is no silver bullet. Precisely defining the problem is always the hard part. And the best way to precisely define a problem and its solution is code.
I’ll let other people fight swarms of bots building… well who knows what. Maybe someday it will deliver useful stuff, but I’m highly skeptical.
Much of it is just "put this magic string before your prompt to make the LLM 10x better" voodoo, similar to the SEO voodoo common in the 2000s.
just remember that it works the same for everyone: you input text, magic happens, text comes out.
if you can properly explain a software engineering problem in plain language, you're an expert in using LLMs. everything on top of that people experimenting or trying to build the next big thing.
Maybe this is a naive question, but how are "skills" different from just adding a bunch od examples of good/bad behavior into the prompt? As far as I can tell, each skill file is a bunch of good/bad examples of something. Is the difference that the model chooses when to load a certain skill into context?
This style of prompting, where you set up a dire scenario in order to try to evoke some "emotional" response from the agent, is already dated. At some point, putting words like IMPORTANT in all uppercase had some measurable impact, but at the present time, models just follow instructions.
Save yourself the experience of having to write and maintain prompts like this.
What’s irritating is that the llms haven’t learned this as bout themselves yet. If you ask an llm to improve its instructions those sort of improvements are what it will suggest.
It is the thing I find most irritating about working with llms and agents. They seem forever a generation behind in capabilities that are self referential.
The past few years have taught me that these are the people that rise to the top of society (much to my chagrin).
The average person doesn’t want to hear from thoughtful intellectuals presenting nuanced opinions. They want to hear from those who brashly and boastfully present themselves as authority figures, and then bolster the listeners preconceived ideas with violently exaggerated language. Shallow but sensational is what sells.
I think that Elons bombastic claims about self driving have really popularized this approach. But you can now see it everywhere in tech: bitcoin going to $1B and nocoiners will be peasants, AI is going to turn us all in to paperclips, and on and on…
I'm far from an AI enthusiast but I really appreciate Simon for his articles and takes on AI. He's enthusiastic and optimistic but that doesn't make him a hype man.
It likely works - but knowing that YAGNI is a thing, means at some level you are invoking a cultural touchstone for a very specific group of humans.
Edit -
I dug into the superpowers and skills for a bit. Definitely learned from it.
There’s stuff that doesn’t make sense to me on a conceptual basis. For example in the skill to preserve productive tensions. There’s a part that goes :
> The trade-off is real and won't disappear with clever engineering
There’s no dimension for “valid” or prediction for tradeoff.
I can guess that if the preceding context already outlines tradeoffs clearly, or somehow encodes that there is no clever solution that threads the needle - then this section can work.
Just imagining what dimensions must be encoding some of this suggests that it’s … it won’t work for situations where the example wasn’t already encoded in the training. (Not sure how to phrase it)
> some of the ones I've played with come from telling Claude "Here's my copy of programming book. Please read the book and pull out reusable skills that weren't obvious to you before you started reading
This is actually a really cool idea. I think a lot of the good scaffolding right now is things like “use TDD” bit if you link citations to the book, then it can perhaps extract more relevant wisdom and context (just like I would by reading the book), weather than using the generic averaged interpretation of TDD derived from the internet.
I do like the idea of giving your Claude a reading list and some spare tokens on the weekend where you’re not working, and having it explore new ideas and techniques to bring back to your common CLAUDE.md.
> It also bakes in the brainstorm -> plan -> implement workflow I've already written about. The biggest change is that you no longer need to run a command or paste in a prompt. If Claude thinks you're trying to start a project or task, it should default into talking through a plan with you before it starts down the path of implementation.
... So, we're refactoring the process of prompting?
> As Claude and I build new skills, one of the things I ask it to do is to "test" the skills on a set of subagents to ensure that the skills were comprehensible, complete, and that the subagents would comply with them. (Claude now thinks of this as TDD for skills and uses its RED/GREEN TDD skill as part of the skill creation skill.)
> The first time we played this game, Claude told me that the subagents had gotten a perfect score. After a bit of prodding, I discovered that Claude was quizzing the subagents like they were on a gameshow. This was less than useful. I asked to switch to realistic scenarios that put pressure on the agents, to better simulate what they might actually do.
... and debugging it?
... How many other basic techniques of SWEng will be rediscovered for the English programming language?
documents like https://github.com/obra/superpowers/blob/main/skills/testing... are very confusing to read as a human.
"skills" in this project generally don't seem to follow set format and just look like what you would get when prompting an LLM to "write a markdown doc that step by step describes how to do X" (which is what actually happened according to the blog post).
idk, but if you already assume that the LLM knows what TDD is (it probably ingested ~100 whole books about it), why are we feeding a short (and imo confusing) version of that back to it before the actual prompt?
i feel like a lot of projects like this that are supposed to give LLMs "superpowers" or whatever by prompt engineering are operating on the wrong assumption that LLMs are self-learning and can be made 10x smarter just by adding a bit of magic text that the LLM itself produced before the actual prompt.
ofc context matters and if i have a repetitive tasks, i write down my constraints and requirements and paste that in before every prompt that fits this task. but that's just part of the specific context of what i'm trying to do. it's not giving the LLM superpowers, it's just providing context.
i've read a few posts like this now, but what i am always missing is actual examples of how it produces objectively better results compared to just prompting without the whole "you have skill X" thing.
I fully agree. I’ve been running codex with GPT Pro (5o-codex-high) for a few weeks now, and it really just boils down to context.
I’ve found the most helpful things for me is just voice to Whisper to LLMs, managing token usage effectively and restarting chats when necessary, and giving it quantified ways to check when its work is done (say, AI-Unit-Tests with apis or playwright tests.) Also, every file I own is markdown haha.
And obviously having different AI chats for specialized tasks (the way the math works on these models makes this have much better results!)
All of this has allowed me to still be in the PM role like he said, but without burning down a needless forest on having it reevaluate things in its training set lol. But why would we go back to vendor lock in with Claude? Not to mention how much more powerful 5o-codex-high is, it’s not even close
The good thing about what he said is getting AI to work with AI, I have found this to be incredibly useful in promoting, and segmenting out roles
Everything is just context, of course. Every time I see a blog post on "the nine types of agentic memory" or some such I have a similar reaction.
I would say that systems like this are about getting the agent to correctly choose the precisely correct context snippet for the exact subtask it's doing at a given point within a larger workflow. Obviously you could also do that manually, but that doesn't scale to running many agents in parallel, or running automomously for longer durations.
Also the format seems quite badly written. Ie. those “quick references” are actually examples. Several generic sentences are repeated multiple times in different wording across sections, etc.
This article left me wishing it was "How I'm using coding agents to do <x> task better"
I've been exploring AI for two years now. It's certainly upgraded itself from the toy classification to a basic utility. However, I increasingly run into its limitations and find reverting to pre-LLM ways of working more robust, faster, and more mentally sustainable.
Does someone have concrete examples of integrating LLM in a workflow that pushes state-of-the-art development practices & value creation further?
The XDG spec has been out for decades now. Why are new applications still polluting my HOME? Also seems weird that real data would be put under a cache/ location, but whatever.
It's in the cache location because it's a copy of a plugin that was installed from a GitHub repository, so that's not the original point of truth for that file.
The post reads like the someone throwing bones and reading their fortune. That part where Claude did its own journaling was so cringe it was hilarious. The tone of the journal entry was exactly like the blog author, which suggests to me Claude is reflecting back what the author wants to hear. I feel like Jesse is consumed in a tornado of llm sycophancy.
66 comments
[ 3.9 ms ] story [ 83.0 ms ] thread(There probably are some. Most likely I notice the bad ones more than the good ones. But it does seem like I notice a lot of bad ones, and never any good ones.)
[EDITED to add:] For context, the actual article title begins "Superpowers: How I'm using ..." and it has been auto-rewritten to "Superpowers: I'm using ...", which completely changes what "Superpowers" is understood as applying to. (The actual intention: superpowers for LLM coding agents. The meaning after the change: LLM coding agents as superpowers for humans.)
According to this, we'll all be reading the feelings journals of our LLM children and scolding them for cheating on our carefully crafted exams instead of, you know, making things. We'll read psychology books, apparently.
I like reading and tinkering directly. If this is real, the field is going to leave that behind.
Is Claude really "learning new skills" when you feed it a book, or does it present it like that because you're prompting encourages that sort of response-behavior. I feel like it has to demo Claude with the new skills and Claude without.
Maybe I'm a curmudgeon but most of these types of blogs feel like marketing pieces with the important bit is that so much is left unsaid and not shown, that it comes off like a kid trying to hype up their own work without the benefit of nuance or depth.
maybe explicit support from providers would make it feasible?
I don’t like the looks of that. If I used this, how soon before those instructions would be in conflict with my actual priorities?
Not everything can be the first law.
Spend some time digging around in his https://github.com/obra/Superpowers repo.
I wrote some notes on this last night: https://simonwillison.net/2025/Oct/10/superpowers/
Not only do I have know everything about the code, data and domain, but now I need to understand this whole AI system which is a meta skill of its own.
I fear I may never be able catch up till someone comes along and simplifies it for pleb consumption.
No matter what you are told, there is no silver bullet. Precisely defining the problem is always the hard part. And the best way to precisely define a problem and its solution is code.
I’ll let other people fight swarms of bots building… well who knows what. Maybe someday it will deliver useful stuff, but I’m highly skeptical.
just remember that it works the same for everyone: you input text, magic happens, text comes out.
if you can properly explain a software engineering problem in plain language, you're an expert in using LLMs. everything on top of that people experimenting or trying to build the next big thing.
Just give it a few months. If some technics really work, it’ll get streamlined.
Save yourself the experience of having to write and maintain prompts like this.
That paper is about using persuasion prompts to overcome trained in "safety" refusals, not to improve prompt conformance.
It is the thing I find most irritating about working with llms and agents. They seem forever a generation behind in capabilities that are self referential.
The average person doesn’t want to hear from thoughtful intellectuals presenting nuanced opinions. They want to hear from those who brashly and boastfully present themselves as authority figures, and then bolster the listeners preconceived ideas with violently exaggerated language. Shallow but sensational is what sells.
I think that Elons bombastic claims about self driving have really popularized this approach. But you can now see it everywhere in tech: bitcoin going to $1B and nocoiners will be peasants, AI is going to turn us all in to paperclips, and on and on…
https://news.ycombinator.com/newsguidelines.html
(We detached this subthread from https://news.ycombinator.com/item?id=45549522.)
Edit: your account has unfortunately been doing this repeatedly (https://news.ycombinator.com/item?id=45551198), and you've been breaking the site guidelines in other ways as well (e.g. https://news.ycombinator.com/item?id=45527456). We ban accounts that do this, so if you wouldn't mind reviewing https://news.ycombinator.com/newsguidelines.html and taking the intended spirit of the site more to heart, that would be good.
Is this just someone who has tingly feelings about Claude reiterating stuff back to them? cuz that's what an LLM does/can do
This is voodoo.
It likely works - but knowing that YAGNI is a thing, means at some level you are invoking a cultural touchstone for a very specific group of humans.
Edit -
I dug into the superpowers and skills for a bit. Definitely learned from it.
There’s stuff that doesn’t make sense to me on a conceptual basis. For example in the skill to preserve productive tensions. There’s a part that goes :
> The trade-off is real and won't disappear with clever engineering
There’s no dimension for “valid” or prediction for tradeoff.
I can guess that if the preceding context already outlines tradeoffs clearly, or somehow encodes that there is no clever solution that threads the needle - then this section can work.
Just imagining what dimensions must be encoding some of this suggests that it’s … it won’t work for situations where the example wasn’t already encoded in the training. (Not sure how to phrase it)
This is actually a really cool idea. I think a lot of the good scaffolding right now is things like “use TDD” bit if you link citations to the book, then it can perhaps extract more relevant wisdom and context (just like I would by reading the book), weather than using the generic averaged interpretation of TDD derived from the internet.
I do like the idea of giving your Claude a reading list and some spare tokens on the weekend where you’re not working, and having it explore new ideas and techniques to bring back to your common CLAUDE.md.
... So, we're refactoring the process of prompting?
> As Claude and I build new skills, one of the things I ask it to do is to "test" the skills on a set of subagents to ensure that the skills were comprehensible, complete, and that the subagents would comply with them. (Claude now thinks of this as TDD for skills and uses its RED/GREEN TDD skill as part of the skill creation skill.)
> The first time we played this game, Claude told me that the subagents had gotten a perfect score. After a bit of prodding, I discovered that Claude was quizzing the subagents like they were on a gameshow. This was less than useful. I asked to switch to realistic scenarios that put pressure on the agents, to better simulate what they might actually do.
... and debugging it?
... How many other basic techniques of SWEng will be rediscovered for the English programming language?
idk, but if you already assume that the LLM knows what TDD is (it probably ingested ~100 whole books about it), why are we feeding a short (and imo confusing) version of that back to it before the actual prompt?
i feel like a lot of projects like this that are supposed to give LLMs "superpowers" or whatever by prompt engineering are operating on the wrong assumption that LLMs are self-learning and can be made 10x smarter just by adding a bit of magic text that the LLM itself produced before the actual prompt.
ofc context matters and if i have a repetitive tasks, i write down my constraints and requirements and paste that in before every prompt that fits this task. but that's just part of the specific context of what i'm trying to do. it's not giving the LLM superpowers, it's just providing context.
i've read a few posts like this now, but what i am always missing is actual examples of how it produces objectively better results compared to just prompting without the whole "you have skill X" thing.
I’ve found the most helpful things for me is just voice to Whisper to LLMs, managing token usage effectively and restarting chats when necessary, and giving it quantified ways to check when its work is done (say, AI-Unit-Tests with apis or playwright tests.) Also, every file I own is markdown haha.
And obviously having different AI chats for specialized tasks (the way the math works on these models makes this have much better results!)
All of this has allowed me to still be in the PM role like he said, but without burning down a needless forest on having it reevaluate things in its training set lol. But why would we go back to vendor lock in with Claude? Not to mention how much more powerful 5o-codex-high is, it’s not even close
The good thing about what he said is getting AI to work with AI, I have found this to be incredibly useful in promoting, and segmenting out roles
I would say that systems like this are about getting the agent to correctly choose the precisely correct context snippet for the exact subtask it's doing at a given point within a larger workflow. Obviously you could also do that manually, but that doesn't scale to running many agents in parallel, or running automomously for longer durations.
I've been exploring AI for two years now. It's certainly upgraded itself from the toy classification to a basic utility. However, I increasingly run into its limitations and find reverting to pre-LLM ways of working more robust, faster, and more mentally sustainable.
Does someone have concrete examples of integrating LLM in a workflow that pushes state-of-the-art development practices & value creation further?