I’m curious how you think this compares to the Research -> Plan -> Implement method and prompts from the “Advanced Context Engineering from Agents” video when it comes to actual coding performance on large codebases. I think picking up skills is useful for broadening agents abilities, but I’m not sure I’d that’s the right thing for actual development.

The packaged collection is very cool and so is the idea of automatically adding new abilities, but I’m not fully convinced that this concept of skills is that much better than having custom commands+sub-agents. I’ll have to play around with it these next few days and compare.

Using Research->Plan->Implement flow is orthogonal, though I notice parts of those do exist as skills too. But you sometimes need to do other things too, e.g. debugging in the course of implementing or specific techniquws to improve brainstorming/researching.

Some of these skills are probably better as programmed workflows that the LLM is forced to go through to improve reliability/consistency, that's what I've found in my own agents, rather than using English to guide the LLM and trusting it to follow the prescribed set of steps needed. Some mix of LLMs (choosing skills, executing the fuzzy parts of them) and just plain code (orchestration of skills) seems like the best bet to me and what I'm pursuing.

Orthogonal means there should not be any overlap.

Yeah, but you knew what he meant.

simon is the biggest ai hype man :) . everything is always strong, wild , superb , amazing. his blog posts like the one he linked here never explain why it is wild , strong ,super or any of other superlatives he uses.

also whats with overuse of these words in ai space. every team is 'cracked' , everything is 'wild'. real car salesman vibes.

Here's a counter-example for you from the another day: https://simonwillison.net/2025/Oct/8/claude-datasette-plugin...

> This isn’t necessarily surprising, but it’s worth noting anyway. Claude Sonnet 4.5 is capable of building a full Datasette plugin now.

I do worry a bit about how often I use positive adjectives. If something isn't notable I won't write about it though. In this particle case Jesse's prompting / skills stuff really does deserve the superlatives IMO.

well explain why OPost is "wild" and what makes you recommend it "strongly" .

what have u built with to come to those conclusions ? is this too much to ask.

I recommend it strongly because the "skills" mechanism it describes is a new and very promising technique, and this is the best article I've seen that explains that.

It's "wild" because, among many other experiments, Jesse has experimented with giving Claude a "feelings journal" and prompting it using Graphviz DOT diagrams.

For my previous writing and work on this you can consult my blog - here's the AI-assisted programming tag: https://simonwillison.net/tags/ai-assisted-programming/

I'll admit I struggle to isolate what's fundamentally different between these skill configurations and the usual CLAUDE/AGENTS/etc.md. Clearly there's something; I'm just curious what's the mechanism at play, precisely.

EDIT: Simon actually already answered that point here: https://news.ycombinator.com/item?id=45550115

[flagged]

I read a lot of articles about tricks for using coding agents. I link to very few of them because they don't often present anything new.

Jesse presented something new.

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 try very hard to fulfill the role of "thoughtful intellectuals presenting nuanced opinions" in the AI space. I'm disappointed that I've failed to meet that standard in your eyes.

it was fine to say this is something new and ppl should maybe look at it. but strongly recommending its and calling it wild is not a nuanced opinion. esp given you have not actually used it yourself in any serious way. is it really that big a ask to only use hype words for things you've actually used and made a big difference to you.

edit: looks like my comments crossed into personally attacking you . i apologize for it.

I stand firmly by "wild", by which I mean unexpected, surprising and unconventional:

- Giving Claude a skill that creates more skills is recursive and unconventional

- "Please read the book and pull out reusable skills that weren't obvious to you before you started reading" is a creatively weird and ambitious thing to do

- "It made sense to me that the persuasion principles I learned in Robert Cialdini's Influence would work when applied to LLMs" - that's a wild thing to say

- How is the concept of a "feelings journal" not wild?

- The original installation method for this (before plugins came along) was to prompt "Hey Claude. Please read https://raw.githubusercontent.com/obra/Superpowers/refs/head... and do what it says"

I'm actually pretty careful about the words I use, and I'm confident "wild" is justified in this case.

That's far from true. Also, please don't cross into personal attack on this site.

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.

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.

This looks like usage rules in Elixir, but for agent behaviors, and currently specifically for Claude: https://hexdocs.pm/usage_rules/readme.html

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.

In my opinion, subagents (or more generally, "agents as tools" as a pattern) are an order-of-magnitude level feature. Soon every CLI agent will have them as a first-class feature (you can get them via custom scripting right now with and CLI agent, albeit less ergonomically).

The ability to isolate context-noisy subtasks (like agentically searching through a large codebase by grepping through dozens of irrelevant files to find the one you actually need) unlocks much longer-running loops, and therefore much more complex tasks.

And you don't need a system this complicated to take advantage of it. Literally just a simple "codebase-searcher" agent (and Claude can vibe the agent definition for you) is enough to see the benefit first-hand. Once you see it, if you're like me, you will see opportunities for subagents everywhere.

This is a great answer IMO.

I think they're worthwhile only as a token context management tool - to complete side quests without using up tokens in your main agent loop.

Using them in a way that doesn't waste tokens is something I haven't fully figured out out yet!

I'm not sure I get this. If anything, they'll consume less tokens, because their context will possibly contain a subset of the original single agent prompt, and they only need to see a subset of the original single agent history.

What am I missing?

Take a look at my example here - having a bunch of sub-agents perform a task consumed 50,000+ tokens each across 5 subtasks, because each one had to consume duplicate information. https://simonwillison.net/2025/Oct/11/sub-agents/

But that's down to the way Claude Code has implemented it? If I code this myself I could engineer so that the subagents don't have overlapping context with the orchestrator.

Also, memory itself can be a tool the subagent calls to retrieve only the stuff it needs.