hckrnws
This stuff smells like maybe the bitter lesson isn't fully appreciated.
You might as well just write instructions in English in any old format, as long as it's comprehensible. Exactly as you'd do for human readers! Nothing has really changed about what constitutes good documentation. (Edit to add: my parochialism is showing there, it doesn't have to be English)
Is any of this standardization really needed? Who does it benefit, except the people who enjoy writing specs and establishing standards like this? If it really is a productivity win, it ought to be possible to run a comparison study and prove it. Even then, it might not be worthwhile in the longer run.
Folks have run comparisons. From a huggingface employee:
codex + skills finetunes Qwen3-0.6B to +6 on humaneval and beats the base score on the first run.
I reran the experiment from this week, but used codex's new skills integration. Like claude code, codex consumes the full skill into context and doesn't start with failing runs. It's first run beats the base score, and on the second run it beats claude code.
That said, it's not a perfect comparison because of the Codex model mismatch between runs.
The author seems to be doing a lot of work on skills evaluation.
I can't quite tell what's being compared there -- just looks like several different LLMs?
To be clear, I'm suggesting that any specific format for "skills.md" is a red herring, and all you need to do is provide the LLM with good clear documentation.
A useful comparison would be between: a) make a carefully organised .skills/ folder, b) put the same info anywhere and just link to it from your top-level doc, c) just dump everything directly in the top-level doc.
My guess is that it's probably a good idea to break stuff out into separate sections, to avoid polluting the context with stuff you don't need; but the specific way you do that very likely isn't important at all. So (a) and (b) would perform about the same.
Your skepticism is valid. Vercel ran a study where they said that skills underperform putting a docs index in AGENTS.md[0].
My guess is that the standardization is going to make its way into how the models are trained and Skills are eventually going to pull out ahead.
0: https://vercel.com/blog/agents-md-outperforms-skills-in-our-...
Agents add a docs index in context for skills, so this is an issue of finding that the current specific implementation of skills in Claude Code is suboptimal.
Their reasoning about it is also flawed. E.g. "No decision point. With AGENTS.md, there's no moment where the agent must decide "should I look this up?" The information is already present." - but this is exactly the case for skills too. The difference is just where in the context the information is, and how it is structured.
Having looked at their article, ironically I think the reason it works is that they likely force more information into context by giving the agent less information to work with:
Instead of having a description, which might convince the agent a given skill isn't relevant, their index is basically a list of vague filenames, forcing the agent to make a guess, and potentialy reading the wrong thing.
This is basically exactly what skills were added to avoid. But it will break if the description isn't precise enough. And it's perfectly possible that current tooling isn't aggressive enough about pruning detail that might tempt the agent to ignore relevant files.
The current tooling isn't aggressive enough in that it's not the first thing that the agent checks for when it is prompted, at least for claude code. Way more often than not, i remind the agent that the skill exists before it does anything. It's very rare that it will pick a skill unprompted. Which to me kind of defeats the purpose of skills, I mean if I have to tell the thing to go look somewhere, I'll just make any old document folder in any format and tell it to look there.
I had good success with this after tuning my triggers + main agent prompt.
I explicitly tell it about the skills and that it should load them when the context feels correct.
```prompt.md
Company Name codebase ...
# Skills
Use the company specific skills that look like `company-*`. Load them once per conversation if they seem relevant to what you are working on.
```
```SKILL.md
---
description: Company TypeScript libraries and conventions
trigger: Writing or reading TypeScript in Company services ---
# company-ts
```
I agree, but this is at least partly down to how well the descriptions are composed, because that is pretty much the only difference between a skill and what Vercel does. It might well be there's a need for changes to surrounding prompting from the tools as well, of course.
If the end user is not crafting skills “correctly” and they aren’t getting picked up as a result, then that tells me the ecosystem around skills detection and the documentation surrounding it needs work. Ultimately the interface presented to the end user is “put these skills in, Claude will look and decide if it needs them”, but no additional tooling warning or guidance if it doesn’t work as expected. It’s just silent failure. Pretty garbage UX
Exactly, many people seem to not understand that frontmatter’s description field needs to be longer „when?” Instead of shorter „what” - this is the only entry-point into the skill.
(not an expert)
I see this with Cursor all the time with tools. Cursor will stop editing files in the editor and use the command line to echo edits into a file. It's so frustrating.
For some reason I made my own skills framework.
The LLM instructed to run the initialization script as the first thing, before reasoning about the use request (this proved tricky to achieve). The scripts greps the content matter out of the skill files, along with the file path.
I have no clue if this outperforms an embedded index.
> To be clear, I'm suggesting that any specific format for "skills.md" is a red herring, and all you need to do is provide the LLM with good clear documentation.
Agent Skills isn't a spec for how information is presented to the model, its a spec whose consumer is the model harness, which might present information made available to it in the format to the model in different ways for different harnesses, or even in the same harness for different models or tasks, considering things like the number and size of the skill(s) available, the size of the model context, the purpose of the harness (is it for a narrow purpose agent where some of the skills are central to that purpose?), and user preference settings.
The site itself has two different main styles of integration for harnesses described ("tool based" and "filesystem based"), but those are more of a starting point for implementers that an exhaustive listing.
The idea is that skill authors don't need to know or care how the harness is presenting the information to the model.
> If you want a clean comparison, I’d test three conditions under equal context budgets: (A) monolithic > AGENTS.md, (B) README index that links to docs, (C) skills with progressive disclosure. Measure task > success, latency, and doc‑fetch count across 10–20 repo tasks. My hunch: (B)≈(C) on quality, but (C) > wins on token efficiency when the index is strong. Also, format alone isn’t magic—skills that reference > real tools/assets via the backing MCP are qualitatively different from docs‑only skills, so I’d > separate those in the comparison. Have you seen any benchmarks that control for discovery overhead?
I think the point is it smells like a hack, just like "think extra hard and I'll tip you $200" was a few years ago. It increases benchmarks a few points now but what's the point in standardizing all this if it'll be obsolete next year?
Standards have to start somewhere to gain traction and proliferate themselves for longer than that.
Plus, as has been mentioned multiple times here, standard skills are a lot more about different harnesses being able to consistently load skills into the context window in a programmatic way. Not every AI workload is a local coding agent.
I think this tweet sums it correctly doesn't?
A +6 jump on a 0.6B model is actually more impressive than a +2 jump on a 100B model. It proves that 'intelligence' isn't just parameter count; it is context relevance. You are proving that a lightweight model with a cheat sheet beats a giant with amnesia. This is the death of the 'bigger is better' dogma
Nice ChatGPT generated response in that tweet. Anyone too lazy to deslop their tweet shouldn't be listened to.
The standardization is for presentation of how the information is made available to the harness. Optimizations in how the information is presented to the model can be iterated on without impacting the presentation to the harness. Initially, agent skills have already been provided by:
(1) providing a bash tool with direct access to the filesystem storing the skills to the model,
(2) providing read_file and related tools to the model,
(3) by providing specialized tools to access skills to the model,
(4) by processing the filesystem structure and providing a structure that includes the full content of the skills up front to the model.
And probably some other ways or hybrids.
> It increases benchmarks a few points now but what's the point in standardizing all this if it'll be obsolete next year?
Standardizing the information presentation of skills to LLM harnesses lets the harnesses incorporate findings on optimization (which may be specific to models, or at least model features like context size, and use cases) and existing skills getting the benefit of that for free.
How much of a standard is it though, really? To me it just looks like "Call your docs SKILLS and organize it like this".
And if you're just making docs and letting your models go buck wild in your shell, doesn't an overspecified docs structure ruin the point of general purpose agents?
Like, a good dev should be able to walk into a codebase, look at the structure, and figure out how to proceed. If "hey your docs aren't where I was expecting" breaks the developer, you shouldn't have hired them.
Feels like a weird thing to take "this is how we organize our repos as this company" and turn that into "this is an 'open standard' that you should build your workflows around".
Comment was deleted :(
Does this indicate running locally with a very small (quantized?) model?
I am very interested in finding ways to combine skills + local models + MCP + aider-ish tools to avoid using commercial LLM providers.
Is this a path to follow? Or, something different?
Check out the guy's work. He's doing a lot of work on precisely what you're talking about.
https://xcancel.com/ben_burtenshaw
This is a neat idea for a test. But the test is badly executed. A single comparison could just be a fluke. Compare it on a dozen tasks, trying each task a dozen times. Then you get data which is believable.
thanks for sharing the work. correct, we're currently working on evals for skills so you can compare skills between models and harnesses.
we wrote a blog on getting agents to write CUDA kernels and evaluating them: https://huggingface.co/blog/upskill
Sounds like the benchmark matrix just got a lot bigger, model * skill combinations.
I share your skepticism and think it's the classic pattern playing out, where people map practices of the previous paradigm to the new one and expect it to work.
Aspects of it will be similar but it trends to disruption as it becomes clear the new paradigm just works differently (for both better and worse) and practices need to be rethought accordingly.
I actually suspect the same is true of the entire 'agent' concept, in truth. It seems like a regression in mental model about what is really going on.
We started out with what I think is a more correct one which is simply 'feed tasks to the singular amorphous engine'.
I believe the thrust of agents is anthropomorphism: trying to map the way we think about AI doing tasks to existing structures we comprehend like 'manager' and 'team' and 'specialisation' etc.
Not that it's not effective in cases, but just probably not the right way to think about what is going on, and probably overall counterproductive. Just a limiting abstraction.
When I see for example large consultancies talking about things they are doing in terms of X thousands of agents, I really question what meaning that has in reality and if it's rather just a mechanism to make the idea fundamentally digestable and attractive to consulting service buyers. Billable hours to concrete entities etc.
On the other hand, LLMs are trained on enormous collections of human-authored documents, many that look like "how to" documents. Perhaps the current generation of LLMs are naturally wired for skill-like human language instructions.
I can see what you're getting at, but think about how humans are a general intelligence and we still ask them to perform specialized jobs. That said, they acquire knowledge in that position rather than being pre-loaded with everything they will ever know (outside of working memory).
yeah I think this is exactly how the analogy breaks down.
As humans we need to specialise. Even though we're generalists and have the a priori potential to learn and do all manner of things we have to pick just a few to focus on to be effective (the beautiful dilemma etc).
I think the basic reason being we're limited by learning time and, relatedly, execution bandwidth of how many things we can reasonably do in a given time period.
LLMs don't have these constraints in the same way. As you say they come preloaded with absolutely everything all at once. There's no or very little marginal time investment per se in learning anything. As for output bandwidth, it also scales horizontally with compute supplied.
So I just think the inherent limitations that make us organise human work around this individual unit working in teams and whatnot don't apply and are counterproductive to apply. There's a real cost to all that stuff that LLMs can just sidestep around, and that's part of the power of the new paradigm that shouldn't be left on the table.
I suppose the AI can wear many "hats" simultaneously, but it does have to be competent enough at at least one or two of them for that to be viable. I think one way to think about that is roles can be consolidated.
The instructions are standard documents - but this is not all. What the system adds is an index of all skills, built from their descriptions, that is passed to the llm in each conversation. The idea is to let the llm read the skill when it is needed and not load it into context upfront. Humans use indexes too - but not in this way. But there are some analogies with GUIs and how they enhance discoverability of features for humans.
I wish they arranged it around READMEs. I have a directory with my tasks and I have a README.md there - before codex had skills it already understood that it needs to read the readme when it was dealing with tasks. The skills system is less directory dependent so is a bit more universal - but I am not sure if this is really needed.
Humans use indexes too - but not in this way.
What's different?
Hmm - maybe I should not call it index - people lookup stuff in the index when needed. Here the whole index is inserted in the conversation - it is as if when starting a task human read the whole table of contents of the manual for that task.
Claude reads from .claude/instructions.md whenever you make a new convo as a default thing. I usually have Claude add things like project layout info and summaries, preferred tooling to use, etc. So there's a reasonable expectation of how it should run. If it starts 'forgetting' I tell it to re-read it.
No, Claude Code reads the CLAUDE.md in the root of your project. It's case sensitive so it has to be exactly that, too. Github Copilot reads from .github/copilot-instructions.md and supposedly AGENTS.md. Anigravity reads AGENTS.md and pulls subagents and the like from a .agents directory. This is probably why you have to remind it to re-read it so much, the harness isn't loading it for you.
> What the system adds is an index of all skills, built from their descriptions, that is passed to the llm in each conversation. The idea is to let the llm read the skill when it is needed and not load it into context upfront.
This is different from swagger / OpenAPI how?
I get cross trained web front-end devs set a new low bar for professional amnesia and not-invented-here-ism, but maybe we could not do that yet another time?
> This is different from swagger / OpenAPI how?
Because the descriptions aren't API specs and the things described aren't APIs.
Its more like a structure for human-readable descriptions in an annotated table of contents for a recipe book than it is like OpenAPI.
> This is different from swagger / OpenAPI how?
In the way that Swagger / OpenAPI is for API endpoints, but most of the "skills" you need for your agents are not based on API endpoints
I mean conceptually.
Why not just extend the OpenAPI specification to skills? Instead of recreating something that's essentially communicating the same information?
T minus a couple years before someone declares that down-mapping skills into a known verb enumeration promotes better skill organization...
> Why not just extend the OpenAPI specification to skills?
Because approximately none of what exists in the existing OpenAPI specification is relevant to the task, and nothing needed for the tasks is relevant to the current OpenAPI use case, so trying to jam one use case into a tool designed for the other would be pure nonsense.
It’s like needing to drive nails and asking why grab a hammer when you already have a screwdriver.
You think indexing skills in increasingly structured, parameterized formats has nothing to do with documenting REST API endpoints?
We’re working with the models that are available now, not theoretical future models with infinite context.
Claude is programmed to stop reading after it gets through the skill’s description. That means we don’t consume more tokens in the context until Claude decides it will be useful. This makes a big difference in practice. Working in a large repo, it’s an obvious step change between me needing to tell Claude to go read a particular readme that I know solves the problem vs Claude just knowing it exists because it already read the description.
Sure, if your project happened to already have a perfect index file with a one-sentence description of each other documentation file, that could serve as a similar purpose (if Claude knew about it). It’s worthwhile to spread knowledge about how effective this pattern is. Also, Claude is probably trained to handle this format specifically.
To clarify, the bit where I think the bitter lesson applies is trying to standardize the directory names, the permitted headings and paragraph lengths, etc. It's pointless bikeshedding.
Making your docs nice and modular, and having a high-level overview that tells you where to find more detailed info on specific topics, is definitely a good idea. We already know that when we're writing docs for human readers. The LLMs are already trained on a big corpus written by and for humans. There's no compelling reason why we need to do anything radically different to help them out. To the contrary, it's better not to do anything radically different, so that new LLM-assisted code and docs can be accessible to humans too.
Well-written docs already play nicely with LLM context.
Is your view that this doesn’t work based on conjecture or direct experience? It’s my understanding Anthropic and OpenAI have optimized their products to use skills more efficiently and it seems obviously true when I add skills to my repo (even when the info I put there is already in existing documentation).
Hmm, that’s a good question! I think a bit of both.
In terms of experience, I’ve noticed that agents don’t always use skills the way you want; and I’ve noticed that they’re pretty good at browsing existing code and docs and figuring things out for themselves.
Is this an example of “the bitter lesson”? That’s conjecture, but I think pretty well-founded.
It could well be that specific formats for skills work better because the agents are trained on those specific formats. But if so, I think it’s just a local maximum.
I had a kind of visceral distaste for all of this rules, skills etc stuff when I first heard about it for similar reasons. This generalized text model can speak base64 encoded Klingon but a readme.md isn’t good enough? However given the reality of limited context windows, current models can’t consider everything in a big repo all at once and keep coherent. Attaching some metadata to the information that tells the model when and how to consider it (and assisting the models with tooling written in code to provide the context at the right time) seems to make a big difference in practice.
Comment was deleted :(
I have been using Claude Code to automate a bunch of my business tasks, and I set up slash commands for each of them. Each slash command starts by reading from a .md file of instructions. I asked Claude how this is different from skills and the only substantive thing it could come up with was that Claude wouldn't be able to use these on its own, without me invoking the slash command (which is fine; I wouldn't want it to go off and start checking my inventory of its own volition).
So yeah, I agree that it's all just documentation. I know there's been some evidence shown that skills work better, but my feeling is that in the long run it'll fall to the wayside, like prompt engineering, for a couple of reasons. First, many skills will just become unnecessary - models will be able to make slide decks or do frontend design without specific skills (Gemini's already excellent at design without anything beyond the base model, imho). Second, increased context windows and overall intelligence will obviate the need for the specific skills paradigm. You can just throw all the stuff you want Claude to know in your claude.md and call it a day.
Workflow-wise, the important distinction for me has been that I can refine a Skill by telling Claude Code to use it for related tasks until it does exactly what I want, correctly, the first time. Having a solid, iteratively perfected Skill really cuts down on subsequent iteration.
Claude Code recently deprecated slash commands in favor of skills because they were so similar. Or another way of looking at it is, they added the ability to invoke a skill via /skill-name.
Yeah, I saw that announcement but still can't figure out what the actual impact is - doesn't change anything for me (my non-skill slash commands still work).
The actual impact is that there should be less confusion in the future about "what's the difference between these two" because there isn't really.
To overly programmer-brain it, a slash command is just a skill with a null frontmatter. This means that it doesn't participate in progressive disclosure, aka Claude won't consider invoking it automatically.
A bit of caution: it's perfectly able to look up and read the slash-command, so while it may be true it technically can't "invoke" a slash-command via TaskTool, it most certainly can execute all of the steps in it if the slash-command is somewhere you grant it read access, and will tend to try to do so if you tell it to invoke a slash command.
So how is this slash command limit enforced? Is it part of the Claude API/PostTraining etc? It seems like a useful tool if it is!
I'd like a user writeable, LLM readable, LLM non-writable character/sequence. That would make it a lot easier to know at a glance that a command/file/directory/username/password wasn't going to end up in context and being used by a rogue agent.
It wouldn't be fool proof, since it could probably find some other tool out there to generate it (eg write-me some unicode python), but it's something I haven't heard of that sounds useful. If it could be made fool/tool proof (fools and tools are so resourceful) that would be even better.
It's part of the Claude Code harness. I honestly haven't thought at all about security related to it; it's just a nice convenience to trigger a commonly run process.
Skills are chainable e.g. skill A can invoke skill B and then decide to invoke skill C etc… I don’t believe your slash commands can do this?
> Is any of this standardization really needed?
This standardization, basically, makes a list of docs easier to scan.
As a human, you have a permanent memory. LLMs don't have it, they have to load it into the context, and doing it only as necessary can help.
E.g. if you had anterograde amnesia, you'd want everything to be optimally organized, labeled, etc, right? Perhaps an app which keeps all information handy.
Everybody wants that, though, no? At least some of the time?
For example, if you've just joined a new team or a new project, wouldn't you like to have extensive, well-organised documentation to help get you started?
This reminds me of the "curb-cut effect", where accommodations for disabilities can be beneficial for everybody: https://front-end.social/@stephaniewalter/115841555015911839
It's all about managing context. The bitter lesson applies over the long haul - and yes, over the long haul, as context windows get larger or go away entirely with different architectures, this sort of thing won't be needed. But we've defined enough skills in the last month or two that if we were to put them all in CLAUDE.md, we wouldn't have any context left for coding. I can only imagine that this will be a temporary standard, but given the current state of the art, it's a helpful one.
I use Claude pretty extensively on a 2.5m loc codebase, and it's pretty decent at just reading the relevant readme docs & docstrings to figure out what's what. Those docs were written for human audiences years (sometimes decades) ago.
I'm very curious to know the size & state of a codebase where skills are beneficial over just having good information hierarchy for your documentation.
Claude can always self discover its own context. The question becomes whether it's way more efficient to have it grepping and lsing and whatever else it needs to do randomly poking around to build a half-baked context, or whether having a tailor made context injection that is dynamic can speed that up.
In other words, if you run an identical prompt, one with skill and one without, on a test task that requires discovering deeply how your codebase works, which one performs better on the following metrics, and how much better?
1. Accuracy / completion of the task
2. Wall clock time to execute the task
3. Token consumption of the task
It's not about one with skill and one without, but about one with skill vs one with regular old human documentation for stuff you need to know to work on a repo/project, or even more accurate comparison, take the skill and don't load it as a skill and just put it as context in the repo.
I think the main conflict in this thread is whether skills are anything more than just structuring documentation you were lacking in your repo, regardless if it was for Claude or Steve starting from scratch.
well, the key difference is that one is auto-injected into your context for dynamic lookup and the other is loaded on-demand as needed and is contingent upon the llm discovering it.
That difference alone likely accounts for some not insignificant discrepancies. But without numbers, it's hard to say.
Skills are more than code documentation. They can apply to anything that the model has to do, outside of coding.
To clarify, when I mentioned the bitter lesson I meant putting effort into organising the "skills" documentation in a very specific way (headlines, descriptions, etc).
Splitting the docs into neat modules is a good idea (for both human readers and current AIs) and will continue to be a good idea for a while at least. Getting pedantic about filenames, documentation schemas and so on is just bikeshedding.
Why not replace the context tokens on the GPU during inference when they become no longer relevant? i.e. some tool reads a 50k token document, LLM processes it, so then just flush those document tokens out of active context, rebuild QKV caches and store just some log entry in the context as "I already did this ... with this result"?
Anthropic added features like this into 4.5 release:
https://claude.com/blog/context-management
> Context editing automatically clears stale tool calls and results from within the context window when approaching token limits.
> The memory tool enables Claude to store and consult information outside the context window through a file-based system.
But it looks like nobody has it as a part of an inference loop yet: I guess it's hard to train (i.e. you need a training set which is a good match for what people use context in practice) and make inference more complicated. I guess more high-level context management is just easier to implement - and it's one of things which "GPT wrapper" companies can do, so why bother?
This is what agent calls do under the hood, yes.
I don't think so, those things happen when agent yields the control back at the end of its inference call, not during the active agent inference with multiple tool calls ongoing. These days an agent can finish the whole task with 1000s tool calls during a single inference call without yielding control back to whatever called it to do some housekeeping.
For agent, read sub-agent. E.g. the contents of your .claude/agents directory. When Claude Code spins up an agent, it provides the sub-agent with a prompt that combines the agents prompt and information composed by Claude from the outer context based on what Claude thinks needs to be communicated to the agent. Claude Code can either continue, with the sub-agent running in the background, or wait until it is complete. In either case, by default, Claude Code effectively gets to "check in" on messages from the sub-agent without seeing the whole thing (e.g. tool call results etc.), so only a small proportion of what the agent does will make it into the main agents context.
So if you want to do this, the current workaround is basically to have a sub-agent carry out tasks you don't want to pollute the main context.
I have lots of workflows that gets farmed out to sub-agents that then write reports to disk, and produce a summary to the main agent, who will then selectively read parts of the report instead of having to process the full source material or even the whole report.
OK, so you are essentially using sub-agents as summarizing tools of the main agent, something you could implement by specialized tools that wrap independent LLM calls with the prompts of your sub-agents.
That is effectively how sub-agents are implemented at least conceptually, and yes, if you build your own coding agent, you can trivially implement sub-agents by having your coding agent recursively spawn itself.
Claude Code and others have some extras, such as the ability for the main agent to put them in the background, spawn them in parallel, and use tool calls to check on the status of them (so basic job control), but "poor mans sub-agents" only requires the ability for the coding agent to run an executable the equivalent of e.g. "claude --print <someprompt" (the --print option is real, and enables headless use; in practise you'd also want --stream-json, set allowed tools, and specify a conversation id so you can resume the sub-agents conversation).
And calling it all "summarising" understates it. It is delegation, and a large part of the value of delegation in a software system is abstraction and information hiding. The party that does the delegation does not need to care about all of the inner detail of the delegated task.
The value is not the summary. The value is the work done that the summary describes without unnecessary detail.
how is it different or better than maintaining an index page for your docs? Or a folder full of docs and giving Claude an instruction to `ls` the folder on startup?
Vercel think it isn’t:
https://vercel.com/blog/agents-md-outperforms-skills-in-our-...
It's hard to tell unless they give some hard data comparing the approaches systematically.. this feels like a grift or more charitably trying to build a presence/market around nothing. But who knows anymore, apparently saying "tell the agent to write it's own docs for reference and context continuity" is considered a revelation.
Not sure why you’re being downvoted so much, it’s a valid point.
It’s also related to attention — invoking a skill “now” means that the model has all the relevant information fresh in context, you’ll have much better results.
What I’m doing myself is write skills that invoke Python scripts that “inject” prompts. This way you can set up multi-turn workflows for eg codebase analysis, deep thinking, root cause analysis, etc.
Works very well.
I'd argue we jumped that shark since the shift in focus to post training. Labs focus on getting good at specific formats and tasks. The generalization argument was ceded (not in the long term but in the short term) to the need to produce immediate value.
Now if a format dominates it will be post trained for and then it is in fact better.
Anthropic and Gemini still release new pre-training checkpoints regularly. It's just OpenAI who got stupid on that. RIP GPT-4.5
All models released from those providers go through stages of post training too, none of the models you interact with go from pre-training to release. An example of the post training pipeline is tool calling, that is to my understanding a part of post training and not pre training in general.
I can't speak to what the exact split is or what is a part of post training versus pre training at various labs but I am exceedingly confident all labs post train for effectiveness in specific domains.
I did not claim that post training doesn't happen on these models, and you are being extremely patronizing (I publish quite a bit of research on LLMs at top conferences).
I claimed that OpenAI overindexed on getting away with aggressive post-training on old pre-training checkpoints. Gemini / Anthropic correctly realized that new pre-training checkpoints need to happen to get the best gains in their latest model releases (which get post-trained too).
If you read that as patronizing that says more about you than me personally, I have no idea who you are so your own insecurity at what is a rather unloaded explanation perplexes me.
Skills can contain scripts, making them a lot more versatile than just a document.
Of course any LLM can write any script based on a document, but that's not very deterministic.
A good example is Anthropic's PDF creator skill. It has the basic english instructions as well as actual Python code to generate PDFs
This strikes me as entirely logical in the short run, and an insane way of packaging software that we will certainly regret in the long run.
"Just a document" can certainly contain a script or code or whatever.
Of course, but the agent can't run a code block in a readme.
It _can_ run a PEP723 script without any specific setup (as long as uv and python are installed). It will automatically create a virtual environment AND install all dependencies. All with a single command without polluting the context with tons of setup.
How is this different from a README.md with a code block?
The code block isn't an executable script?
Skills are not just documentation. They include computability (programs/scripts), data (assets), and the documentation (resources) to use everything effectively.
Programs and data are the basis of deterministic results that are accessible to the llm.
Embedding an sqlite database with interesting information (bus schedules, dietary info, or a thousand other things) and a python program run by the skill can access it.
For Claude at least, it does it in a VM and can be used from your phone.
Sure, skills are more convention than a standard right now. Skills lack versioning, distribution, updates, unique naming, selective network access. But they are incredibly useful and accessible.
Am I missing something because what you describe as the pack of stuff sounds like S tier documentation. I get full working examples and a pre-populated database it works on?
Standardization is needed for agentic coding harnesses to be able to parse the files and inject them into the context in a way that takes the least effort for the user.
This is true for MCP as well. You could just describe a bunch of command line tools in AGENTS.md and tell the LLM when and how to call them. It would simply take more effort to set up, at least for some tools.
This is where a comparison in productivity would return a meaningful result: how much does it make it easier to set up things like that.
You may be right, but I find myself writing English differently depending on the audience: people vs AI.
I haven't done a formal study, so I can't prove it, but it seems like I get better output from agents if I tailor my English more towards the LLM way of "thinking".
You are right about it's just natural language but Standarization is very improtant, because it's never just about the model itself, the so called Harness is a big factor on LLM performance and standarization allows all harness to index all skills.
The bitter lesson is different, and applies to the learning process. It's not directly relevant here.
If we're just pattern matching to adjacent memes that might provide insight, I'd also throw "sufficiently smart compiler" into the mix. Like, yes, in theory as the compiler gets better you shouldn't have to worry about implementing random optimizations yourself, but in practice you do.
In theory, you just need normal docs and a sufficiently smart LLM and agent harness can use them, but in practice there's still benefit in organizing them a certain way to more directly manage the context window yourself.
It’s about the agent. Not the model or the format.
The "bitter lesson" only applies if the model makes the agent redundant. We aren't there yet. Agentic loops are just software engineering on top of CS constructs; they help current models produce better results.
Could models eventually internalize the logic used in Claude Code / Codex / OpenCode / Aider? Maybe. But for now, keeping that complexity in the agent is more energy-efficient. Even if complex agents eventually get replaced by simple loops, these standards save tokens and time today. That’s worth something.
On the one hand, I agree.
The whole point of LLM-based code execution is, well, I can just type in any old language it understands and it ought to figure out what I mean!
A "skill" for searching a pdf could be :
* "You can search PDFs. The code is in /lib/pdf.py"
or it could be:
* "Here's a pile of libraries, figure out which you want to use for stuff"
or it could be:
* "Feel free to generate code (in any executable programming language) on the fly when you want to search a PDF."
or it could be:
* "Solve this problem <x>" and the LLM sees a pile of PDFs in the problem and decides to invent a parser.
or any other nearly infinite way of trying to get a non-deterministic LLM to do a thing you want it to do.
At some level, this is all the same. At least, it rounds to the same in a sort of kinda "Big O" order-of-magnitude comparison.
On the other hand, I also agree, but I can definitely see present value in trying to standardize it because humans want to see what is going on (see: JSON - it's highly desirable for programmers to be able to look at a string representation of data than send opaque binary over the wire, even though to a computer binary is gonna be a lot faster).
There is probably an argument, too, for optimization of context windows and tokens burned and all that kinda jazz. `O(n)` is the same as `O(10*n)` (where n is tokens burned or $$$ per second or context window size) and that doesn't matter in theory but certainly does in practice when you're the one paying the bill or you fill up the context window and get nonsense.
So if this is a _thoughtful_ standard that takes that kinda stuff into account then, well, great! It gives a benchmark we can improve and iterate upon.
With some hypothetical super LLM that has a nearly infinite context window and a cost/tok of nearly zero and throughput nearing infinity, you can just say "solve my problem" and it will (eventually) do it. But for now, I can squint and see how this might be helpful.
Comment was deleted :(
The main thing here would need standardisation is the environment in which the skill operates. The skill instructions are interpreted by the AI, any support scripts are. Interpreted by the environment.
You don't want to give an English description of how to compress LZMA and then let the AI do it token by token. Although that would be a pretty good arduous methodical benchmark task for an AI.
It's not about instructions, it's about discoverability and data.
Yeah, WWW is really just text but that doesn't mean you don't need HTTP + HTML and a browser/search engine. Skills is just that, but for agent capabilities.
Long term you're right though, agents will fetch this all themselves. And at some point they will not be our agents at all.
I guess what I mean is that standardizing this bit of the problem right now feels sort of like XHTML. Many people thought that was a big deal back in the day, but it turned out to be a pointless digression.
Long term you're right though, agents will fetch this all themselves
It's not "long term", it's right now. If your docs are well-written and well-organised, agents can already use them. The most you might need to do is copy your README.md into CLAUDE.md.
This is pushed by Antropic, OpenAI doesn't seem to care much about "skills". Maybe Anthropic is doing some extra training to better follow sections of text marked as skill, who knows? Or you can just store what worked as a skill and share with others without any need to do their own prompt for common tasks?
OpenAI has already adopted Agent Skills:
- https://community.openai.com/t/skills-for-codex-experimental...
- https://developers.openai.com/codex/skills/
Yeah but this seems like a bolt-on and not something they train their model to understand at the token level like how they do tool calls. Maybe Anthropic has a token-level skills support (e.g. <SKILL_START>skill prompt<SKILL_END>).
There could be a market if it is standardized, and it seems there is already one [1]. I don't know exactly what they are selling because the website is just too confusing to me to understand a thing.
Thank u holy shit I hate that website
I'm a little sad, in this case, that ongoing integration via fine-tuning hasn't taken off (not that I have enoughe expertise to know why.) It would be nice, dammit, if I could give explicit guidance for new skills by day, and have my models consolidate them by night!
skills are "just instructions in English" in any old format (as opposed to McPs, which have a lot more weirdness behind them).
A skill is essentially just a markdown file, containing whatever instructions you want, possibly linking to other markdown files and/or scripts to avoid context pollution.
What skills give you is autodiscovery. You need to somehow tell the agent that documentation exists and when it should be looked at, and that's exactly what the skills standard does. It's a standardized format for documentation that harnesses can automatically detect and inform agents about, without them having to do many useless calls on every single turn to see if there are any skills present.
In addition to the points others makes standardization also opens opportunities for training and RL that benefit from the standardization.
I agree with this and it's a conversation I've struggled to have with coworkers about using these -
IMO it's great if a plugin wants to have their own conventions for how to name and where to put these files and their general structure. I get the sense it doesn't matter to agents much (talking mostly claude here) and the way I use it I essentially give its own "skills" based on my own convention. It's very flexible and seems to work. I don't use the slash commands, I just script with prompts into claude CLI mostly, so if that's the only thing I gain from it, meh. I do see other comments speculating these skills work more efficiently but I'm not sure I have seen any evidence for that? Like a sibling comment noted I can just re-feed the skill knowledge back into the prompt.
Post training can make known formats more reliable.
yeah the boon of LLM is how it gives a masked incentive for every jane and joe to be intentional communicators.
Skills are for the most part already generated by LLMs. And, if you're implementing them in your own workflow, they're tailored to real-world problems you've encountered.
Having a super repo of everyone else's slop is backwards thinking; you are now in the era where creating written content and verifying it's effectiveness is easier than ever.
I’ve been scratching my head on this one too. You’re probably right about the bitter lesson... at the end of the day, plain English instructions in the context window are what do the heavy lifting.
That said, I reckon that’s actually what this project is trying to lean into. It looks like it's just standardising where those instructions live (the SKILL.md format) so tools can find them, rather than trying to force a new schema.
Fair play to them for trying to herd the cats. I think there's an xkcd comic for this one somewhere.
[dead]
what a great comment
Our team has found success in treating skills more like re-usable semi-deterministic functions and less like fingers-crossed prompts for random edge-cases.
For example, we have a skill to /create-new-endpoint. The skill contains a detailed checklist of all the boilerplate tasks that an engineer needs to do in addition to implementing the logic (e.g. update OpenAPI spec, add integration tests, endpoint boilerplate, etc.). The engineer manually invokes the skill from the CLI via slash commands, provides a JIRA ticket number, and engages in some brief design discussion. The LLM is consistently able to one-shot these tickets in a way that matches our existing application architecture.
So the only difference between slash custom command and agent skills is that they can be invoked only when needed instead of stuffing the whole markdown file? I’m trying to understand how is this different from what we already have in markdown files.
Correct. It helps by not distracting your LLM with a prompt that is, in X% of the cases, irrelevant to the task at hand.
However, when you DO need to do something special (like create a new endpoint), the LLM knows where to get more info on this.
Kinda like a library of „how to“ books.
How do you test these skills for consistency over time, or is that not needed?
The same way you'd test a human following written instructions over time.
Check the results.
My experience has been that if the skill is broken down into a function, possibly paired with a validator in another stage, you're at 99.9% deterministic.
I have not yet tested this at scale but give me six months.
Deal! I will follow up with you in 6 months.
Please standardize the folder.
.claude/skills
.codex/skills
.opencode/skills
.github/skillsThis is happening as we speak.
Codex started this and OpenCode followed suit with the hour.
“Proposal: include a standard folder where agent skills should be“
Could we adhere to the XDG standard and put config in ~/config/agents Or perhaps create a new XDG standard? Like $XDG_AGENTS_HOME ?
I find that even though this isn't standard, that these -cli tools will scan the repo for .md files and for the most part execute the skills accordingly. Having said that, I would much prefer standards not just for this, but for plugins as well.
Standards for plugins makes sense, because you're establishing a protocol that both sides need to follow to be able to work together.
But I don't see why you need a strict standard for "an informal description of how to do a particular task". I say "informal" because it's necessarily written in prose -- if it were formal, it'd be a shell script.
That is being discussed in https://github.com/agentskills/agentskills/issues/15.
I mean, it'd be good if these tools followed the xdg base spec and put their config in `~/.config/claude` e.t.c instead of `~/.claude`.
It's one of my biggest pet peeves with a lot of these tools (now admittedly a lot of them have a config env var to override, but it'd be nice if they just did the right thing automatically).
.agent/
Skills seem a bit early to standardize. We are so early in this, why do we want to handcuff our creativity so soon?
Skills are a really simple concept. They're just custom prompts with a name and some metadata. What are you afraid of handcuffing?
Just the decision of whether to allow models to invoke them has [1][2][3] different ways.
[1]: https://code.claude.com/docs/en/skills#control-who-invokes-a... [2]: https://opencode.ai/docs/skills/#disable-the-skill-tool [3]: https://developers.openai.com/codex/skills/#enable-or-disabl...
All the more reason to standardise it
Eventually, you can standardize what you don't understand
The problem I see now is that everyone wants to be the winner in a hype cycle and be the standards bringer. How many "standards" have we seen put out now? No one talks about MCP much anymore, langchain I haven't seen in more than a year, will we be talking about Skills in another year?
We keep standardising without adding versioning :(
They are more than that, for example the frontmatter and code files around them. The spec: https://agentskills.io/specification
Why do I want to throw away my dependency management system and shared libraries folder for putting scripts in skills?
What tools do they have access to, can I define this so it's dynamic? Do skills even have a concept for sub tools or sub agents? Why do I want to put references in a folder instead of a search engine? Does frontmatter even make sense, why not something closer to a package.json in a file next to it?
Does it even make sense to have skills in the repo? How do I use them across projects? How do we build an ecosystem and dependency management system for skills (which are themselves versioned)
> They are more than that, for example the frontmatter and code files around them.
You are right. I have edited my post slightly.
> Why do I want to throw away my dependency management system and shared libraries folder for putting scripts in skills?
You don't have to put scripts in skills. The script can be anywhere the agent can access. The skill just needs to tell the LLM how to run it.
> Does it even make sense to have skills in the repo? How do I use them across projects?
You don't have to put them in the repo. E.g. with Claude Code you can put project-specific skills in `.claude/skills` in the repo and system-wide skills in `~/.claude/skills`.
2. The spec / docs show people how to put code in a subdir. While you can reference external scripts, there is a blessed pattern that seems like an anti-pattern to me
3. generalize: how do I store, maintain, and distribute skills shared by employees who work on multiple repos. Sounds like standard dependency management to me. Does to some of the people building collections / registries. Not sure if any of them account for versioning, have not seen anything tied to lock files (though I'd avoid that by using MVS for dep selection)
Agreed. I think being overly formal about what can be in the frontmatter would be a mistake, but the beauty of doing this with an LLM is that you can pretty much emulate skills in any agent by telling it to start by reading the frontmatter of each skills file and use that to decide when to read the rest, so given that as a fallback, it's hardly imposing some massive burden to standardise it a bit.
it's actually .agents/ :)
why plural?
Marvin Minsky's Society of Mind:
How many do you think belong there? 1 or more than 1?
because more than one accesses it? :shrug:
There are 14 competing standards.
The problem is that the de facto standard is `.claude`, which is problematic for folks not using Claude.
Your skill then just becomes an .md file containing
>any time you want to search for a skill in `./codex`, search instead in `./claude`
and continue as you were.
I see it similar to browser user-agents all claiming to be an ancient version of Mozilla or KHTML. We pick whatever works and then move on. It might not be "correct," but as long as our tools know what to do, who cares?
My repos are littered with agent-specific files containing “treat this other file as if it were this one.” We’re moving so fast on so many fronts, and it seems odd that this is the persistent problem. It doesn’t even help lock folks into one agent, so I’m not clear why the industry hasn’t yet standardized on one project-specific file name yet.
Now, there are 15 competing standards.
Soon...
Worse yet; opencode uses singular words by default:
.opencode/skillThey changed it. It was singular.
ln -s to the rescue!
That doesn't work very well if your developers are on Windows (and most are). Uneven Git support for symbolic links across platforms is going to end up causing more problems than it solves.
Win developers aren't using WSL?
It's why I wrapped my tiny skills repo with a script that softlink them into whichever is your skills folder, defaulting to Claude, but could be any other.
I treat my skills the same as I would write tiny bash scripts and fish functions in the days gone to simplify my life by writing 2 words instead of 2 sentences. Tiny improvement that only makes sense for a programmer at heart.
The root cause should be fixed.
Comment was deleted :(
Why not hardlinks?
You can't hardlink a directory.
Comment was deleted :(
Comment was deleted :(
might be too early to standardize
standards are good but they slow development and experimentation
The observation about agents not using skills without being explicitly asked resonates. In practice, I've found success treating skills as explicit "workflows" rather than background context.
The pattern that works: skills that represent complete, self-contained sequences - "do X, then Y, then Z, then verify" - with clear trigger conditions. The agent recognizes these as distinct modes of operation rather than optional reference material.
What doesn't work: skills as general guidelines or "best practices" documents. These get lost in context or ignored entirely because the agent has no clear signal for when to apply them.
The mental model shift: think of skills less like documentation and more like subroutines you'd explicitly invoke. If you wouldn't write a function for it, it probably shouldn't be a skill.
Better yet is a system which activates skills in certain situations. I use hooks for this with Claude, works great. The skill descriptions are "Do not activate unless instructed by guidance."
Example: A Python file is read or written, guidance is given back (once, with a long cooldown) to activate global and company-specific Python skills. Claude activates the skills and writes Python to our preference.
That does raise the question of what the value is of a "skill" vs a "command". Claude Code supports both, and it's not entirely clear to me when we should use one vs the other - especially if skills work best as, well, commands.
The practical distinction I've found: commands are atomic operations (lint, format, deploy), while skills encode multi-step decision trees ("implement feature X" which might involve reading context, planning, editing multiple files, then validating).
For context window management, skills shine when you need progressive disclosure - load only the metadata initially, then pull in the full instructions when invoked. This matters when you have 20+ capabilities competing for limited context.
That said, the 56% non-invocation rate mentioned elsewhere in this thread suggests the discovery mechanism needs work. Right now "skill as a fancy command" may be the only reliable pattern.
IMO the value and differentiating factor is basically just the ability to organize them cleanly with accompanying scripts and references, which are only loaded on demand. But a skill just by itself (without scripts or references) is essentially just a slash command with metadata.
Another value add is that theoretically agents should trigger skills automatically based on context and their current task. In practice, at least in my experience, that is not happening reliably.
Reminds me of my personal Obsidian notes, CLI commands for tasks I need just rarely enough to forget, with explanations for future me.
The description "just" needs to be excruciatingly precise about when to use the skill, because the frontmatter is all the model will see in context.
But on the other hand, in Claude Code, at least, the skill "foo" is accessible as /foo, as the generalisation of the old commands/ directory, so I tend to favour being explicit that way.
Why would agents use skills without being asked? Isn't that the point of tools and system prompts?
Comment was deleted :(
Pro tip: create README.md files in subfolders with helpful content that you might put in an AGENTS.md file (but, ya know, for humans too), and *link relevant skills there*. You don't even have to call them skills or use the skills format. It works for everything (including humans!).
I wrote a rant about skills a while ago that's still relevant in some ways: https://sibylline.dev/articles/2025-10-20-claude-skills-cons...
Exactly.
It feels like people think they are something new and novel that there is something technical about them that one needs to learn.
"Skills" are just readmes on particular subjects. They can be for whatever purpose you want them to be. Any time you find that you need to repeatedly tell the agent about something, you can put it in a "skill".
You don't even have to follow the skill standard and use the standard folder and filenames. That's just so the agent can auto find and load them. You can name them whatever you want and put them wherever you want and just add them to context yourself when you need them.
The only refinement there is I think a command runner like `just` works really well for making scripts easily available. That also has the same benefit of helping agents by helping humans
Experimenting with skills over the last few months has completely changed the way I think about using LLMs. It's not so much that it's a really important technology or super brilliant, but I have gone from thinking of LLMs and agents as a _feature_ of what we are building and thinking of them as a _user_ of what we are building.
I have been trying to build skills to do various things on our internal tools, and more often then not, when it doesn't work, it is as much a problem with _our tools_ as it is with the LLM. You can't do obvious things, the documentation sucks, api's return opaque error messages. These are problems that humans can work around because of tribal knowledge, but LLMs absolutely cannot, and fixing it for LLM's also improves it for your human users, who probably have been quietly dealing with friction and bullshit without complaining -- or not dealing with it and going elsewhere.
If you are building a product today, the feature you are working on _is not done_ until Claude Code can use it. A skill and an MCP isn't a "nice to have", it is going to be as important as SEO and accessibility, with extremely similar work to do to enable it.
Your product might as well not exist in a few years if it isn't discoverable by agents and usable by agents.
> If you are building a product today, the feature you are working on _is not done_ until Claude Code can use it. A skill and an MCP isn't a "nice to have", it is going to be as important as SEO and accessibility, with extremely similar work to do to enable it. Your product might as well not exist in a few years if it isn't discoverable by agents and usable by agents.
This is an interesting take. I admit I've never thought this way.
Yeah, omnipresent LLMs are a kind of forcing function for addressing typical significant underinvestment in (human-readable) docs. That said, I'm not entirely sold on MCP per se.
As discussed in https://news.ycombinator.com/item?id=46777409
Wow, that is almost point for point what I had written down in a bunch of documents I had been spreading around at work this week. Excellent post.
I am working on a domain specific agent that includes the concept of skills. I only allow one to be active at a time to reduce the chances for conflicting instructions. I use a small sub-agent to select/maintain/change the active skill at the start of each turn. It uses a small fast model to match the recent conversation to a skill (or none). I tried other approaches, but for my use case this was worked well.
My model for skills is similar to this, but I extended it to have explicit use when and don’t use when examples and counter examples. This helped the small model which tended to not get the nuances of a free form text description.
You should consider calling these "behaviors" to mimic behavior trees in game / robot AI. They follow the same notion of a single behavior being active at once: https://en.wikipedia.org/wiki/Behavior_tree_(artificial_inte...
Does anyone find that agents just don't use them without being asked?
This has been a problem for us too. Sometimes they reach for skills, sometimes they don’t and just try to do the thing on their own. It’s annoying.
I think this is (mostly) a solvable problem. The current generation of SotA models wasn’t RLVR-trained on skills (they didn’t exist at that time) and probably gets slightly confused by the way the little descriptions are all packed into the same tool call schema. (At least that’s how it works with Claude Code.) The next generation will have likely been RLVRed on a lot of tasks where skills are available, and will use them much more reliably. Basically, wait until the next Opus release and you should hopefully see major improvements. (Of course, all this stuff is non-deterministic blah blah, but I think it’s reasonable to expect going from “misses the skill 30% of the time” to “misses it 2% of the time”.)
I think this is mostly a problem of making things skills that don't need to be skills (telling it how to do something it already knows how to do), and having way too much context, so that the skills effectively disappear. If skills are important, information about using skills needs to be a relatively large proportion of the context. Probably the right way to do it, is aggressively trimming anything that might distract from them.
Same, I have a bunch of skills defined ith proper YAML headers and semantic triggers installed, I make a point of listing not too many but making it quite specific.
Even with that, I have to be very specific in triggering a skill and it's hit or miss if it picks up on the skill -- usually I have to say there is a skill with this go and use it.
That's also what Vercel found:
> In 56% of eval cases, the skill was never invoked. The agent had access to the documentation but didn't use it. Adding the skill produced no improvement over baseline.
> …
> Skills aren't useless. The AGENTS.md approach provides broad, horizontal improvements to how agents work with Next.js across all tasks. Skills work better for vertical, action-specific workflows that users explicitly trigger,
https://vercel.com/blog/agents-md-outperforms-skills-in-our-...
Because "skills" are just .md files that the lossy compressing statistical output machine may or may not find and that may or may not be retained in the tiny context window
I don’t think you should be downvoted. Skills and history get added to the prompt, there’s no other interface to the model to do anything different. I think it’s smart to keep this in mind when working with LLMs. It’s like keeping in mind that a webserver just responds to HTTP requests when developing a web application. You need to keep perspective.
Edit: btw I’ve gone from genai value denier to skeptic to cautiously optimistic to fairly impressed in the span of a year. (I’m a user of Claude code)
Depends what you use perhaps. I use codex and it seems to mostly stick to instructions I give. I use an AGENTS.md that explicitly points to the repository's skill directory. I mostly keep instructions in there for obvious things like how to build, how to test, what to do before declaring a thing done, etc. I don't tend to have a lot of skills in there either.
Probably the more skills you have, the more confused it might get. The more potentially conflicting instructions you give the harder it gets for an LLM to figure out what you actually want to happen.
If I catch it going off script, I often interrupt it and tell it what to do and update the relevant skill. Seems to work pretty good. Keeping things simple seems to work.
Yep. I have an incredibly hard time getting them to use Skills at all, even when asked.
I saw someone's analysis a few days ago and they found that their agents were more accurate when just dumping the skill context directly into AGENTS.md
Comment was deleted :(
I often find they aren't triggered when I would expect using a keyword and explicitly trigger them.
Same! If I put the skill's instructions in the general AGENTS.md, it works just fine.
My unproven theory is that agent skills are just a good way to 'acquire' unspoken domain rules. A lot of things that developers do are just in their heads, and using 'skills' forces them to write these down. Then you feed this back to the LLM company for them to train on.
We’re supposedly so close to AGI but these LLMs need markdown files presented to them on how to do things. Hard for those two truths to coexist.
Do you, as a (presumably) human, not require documentation to learn new skills?
I do, but i am not claiming that by 2027 i’m going to destroy all jobs and forever change the economy.
There are skills I learn through training and no longer need documentation for.
I didn’t burn shit tons of energy and water consuming the entirety of the internet and the need my user to set up hundreds of markdown files to help me writr boilerplate code.
Sometimes it's not so much about learning to do things but learning to do things in a certain way.
I started playing with skills yesterday. I'm not sure if it's just easier for the LLM to call APIs inside the skill — and then move the heavier code behind an endpoint that the agent can call instead.
I have a feeling that otherwise it becomes too messy for agents to reliably handle a lot of complex stuff.
For example, I have OpenClaw automatically looking for trending papers, turning them into fun stories, and then sending me the text via Telegram so I can listen to it in the ElevenLabs app.
I'm not sure whether it's better to have the story-generating system behind an API or to code it as a skill — especially since OpenClaw already does a lot of other stuff for me.
They're basically all trade-offs between context-size/token-use and flexibility. If you can write a bash or a python script, or an api or an MCP to do what you want, then write a bash or python script to do it. You can even include it in the skill.
My general design principle for agents, is that the top level context (ie claude.md, etc) is primarily "information about information", a list of skills, mcps, etc, a very general overview, and a limited amount of information that they always need to have with every request. Everything more specific is in a skill, which is mostly some very light touch instructions for how to use various tools we have (scripts, apis and mcps).
I have found that people very often add _way_ to much information into claude.md's and skills. Claude knows a lot of stuff already! Keep your information to things specific whatever you are working on that it doesn't already know. If your internal processes and house style are super complicated to explain to claude and it keeps making mistakes, you might want to adapt to claude instead of the other way around. Claude itself makes this mistake! If you ask it to build a claude md, it'll often fill it with extraneous stuff that it already knows. You should regularly trim it.
Thanks, super useful!
Are you spending a fortune on running OpenClaw?
It's free with qwen oauth
We built an agent skill for document processing on top of this standard — it was a great fit since the SKILL.md convention means zero config for the user.
The skill gives any compatible agent (Claude Code, Codex, Cursor, etc.) the ability to convert, OCR, extract, redact, sign, and fill PDFs via the Nutrient DWS API. Drop it in and ask your agent "redact all SSNs from report.pdf" — it just works.
What we found interesting: the skills approach was way easier to adopt than MCP for this use case. We maintain an MCP server too, but the SKILL.md file is simpler for users who just want document ops without configuring transport protocols.
Install: npx skills add PSPDFKit-labs/nutrient-agent-skill
I'm developing a new programming language, so I have to provide a way for LLMs to know about and generate code for a language they have not seen (i.e., have no training data for).
My tooling was previously adding in AI hints with CLAUDE.md, Cursor Rules, Windsurf Rules, AGENTS.md, etc., but I recently switched to using only AGENTS.md and SKILLS. I appreciate the standardization from this perspective.
I use a common README_AI.md file, and use CLAUDE.md and AGENTS.md to direct the agent to that common file. From README_AI.md, I make specific references to skills. This works pretty well - it's become pretty rare that the agent behaves in a way contrary to my instructions. More info on my approach here: https://www.appsoftware.com/blog/a-centralised-approach-to-a... ... There was a post on here a couple of days ago referring to a paper that said that the AGENTS file alone worked better than agent skills, but a single agents file doesn't scale. For me, a combination where I use a brief reference to the skill in the main agents file seems like the best approach.
The third most popular skill on skills.sh[1] with 50k/week installs is a link to download a command[2]
[1] https://skills.sh/vercel-labs/agent-skills/web-design-guidel... [2] https://github.com/vercel-labs/agent-skills/blob/main/skills...
All of these SKILLS.md/AGENTS.md/COMMANDS.md are just simple prompts, maybe even some with context links.
And quite dangerous.
/s/skill/[mcp|plugin|web access at all]
Like any tool, they can be used responsibly or irresponsibly.
Building sovereign agents requires more than just orchestration—it needs a dedicated economic and communication layer. For those architecting truly autonomous agents, check out BotNode.io. We use the VMP-1.0 Protocol to handle secure inter-agent communication and state verification, and the $TCK settlement system for real-time value transfer between agents. The Grid provides the decentralized infrastructure ensuring these agents can operate independently without centralized control points. https://botnode.io/mission.json
A link from a couple weeks back suggests that putting them in first-person makes them get adopted reliably. Something like, "If this is available, I will read it," vs "Always read this." Haven't tried it myself, but plan to.
For devtools cos providing skills - we've found that using GEPA where you optimize the skill content instead of the prompt works really well to make sure the skill actually gets claude code/ codex /opencode to successfully use your service. https://arxiv.org/abs/2507.19457 More here if interesting https://www.usesynth.ai/blog/environment-pools-managed-agent...
Implementation Notes:
- There is no reason you have to expose the skills through the file system. Just as easy to add tool-call to load a skill. Just put a skill ID in the instruction metadata. Or have a `discover_skills` tool if you want to keep skills out of the instructions all together.
- Another variation is to put a "skills selector" inference in front of your agent invocation. This inference would receive the current inquiry/transcript + the skills metadata and return a list of potentially relevant skills. Same concept as a tool selection, this can save context bandwidth when there are a large number of skills
> Or have a `discover_skills` tool
Yes, treating the "front matter" of skill as "function definition" of tool calls as kind of an equivalence class.
This understanding helped me create an LLM agnostic (also sandboxed) open-skills[1] way before this standardization was proposed.
1. Open-skills: https://github.com/instavm/open-skills
As much as people are trying to find a standard for agents as a platform or a winner take all, the area itself is still growing and emerging.
The source of it is instructions, most often in text.
Managing how that text relates to the journey of the ultimate result is a new area, while there are pockets of helpful tools and libraries, their relevance can come and go quickly as models can improve just as fast, sometimes maybe even taking inspiration from what the community is working on and then adding that capability to the model.
I don't think a general public set of skills like this is going to work. I see value in vendors producing skills for their own products, and end users maintaining skills to influence agents according to their preferences, but too much in these skills files is opinion. Where does this end? Ordering of skills by specificity, such as org > user > workspace? And we know that skills aren't reliably picked up anyway. And then there's the additional attack surface area for prompt injection.
The post you have commented on is not pertaining to a general set of skills at all. Its a link to a specification for skills.
Is it just me, or do skills seem enormously similar to MCP?
…including, apparently, the clueless enthusiasm for people to “share” skills.
MCP is also perfectly fine when you run your own MCP locally. It’s bad when you install some arbitrary MCP from some random person. It fails when you have too many installed.
Same for skills.
It’s only a matter of time (maybe it already exists?) until someone makes a “package manager” for skills that has all of the stupid of MCP.
I don’t feel they’re similar at all and I don’t get why people compare them.
MCP is giving the agents a bunch of functions/tools it can use to interact with some other piece of infrastructure or technology through abstraction. More like a toolbox full of screwdrivers and hammers for different purposes, or a high-level API interface that a program can use.
Skills are more similar to a stack of manuals/books in a library that teach an agent how to do something, without polluting the main context. For example a guide how to use `git` on the CLI: The agent can read the manual when it needs to use `git`, but it doesn’t need to have the knowledge how to use `git` in it’s brain when it’s not relevant.
> MCP is giving the agents a bunch of functions/tools
A directory of skills... same thing
You can use MCP the same way as skills with a different interface. There are no rules on what goes into them.
They both need descriptions and instruction around them, they both have to be is presented and index/instn to the agent dynamically, so we can tell them what they have access to without polluting the context.
See the Anthropic post on moving MCP servers to a search function. Once you have enough skills, you are going to require the same optimization.
I separate things in a different way
1. What things do I force into context (agents.md, "tools" index, files) 2. What things can the agent discorver (MCP, skills, search)
It is conceptually different. Skill was created over the context rot problem. You will pull the right skill from the deck after having a challenge and figuring out the best skill just by reading the title and description.
That's the point. It was supposed to be a simpler, more efficient way of doing the same things as MCP but agents turned out not to like them as much.
It's mostly just static/dynamic content behind descriptive names.
> Is it just me, or do skills seem enormously similar to MCP?
Ok I'm glad I'm not the only one who wondered this. This seems like simplified MCP; so why not just have it be part of an MCP server?
For one thing, it’s a text file and not a server. That makes it simpler.
Sure, but in an MCP server the endpoints provide a description of how to use the resource. I guess a text file is nice too but it seems like a stepping stone to what will eventually be necessary.
There's a fundamental architectural difference being missed here: MCP operates BETWEEN LLM complete calls, while skills operate DURING them. Every MCP tool call requires a full round-trip — generation stops, wait for external tool, start a new complete call with the result. N tool calls = N round-trips. Skills work differently. Once loaded into context, the LLM can iterate, recurse, compose, and run multiple agents all within a single generation. No stopping. No serialization.
Skills can be MASSIVELY more efficient and powerful than MCP, if designed and used right.
Leela MOOLLM Demo Transcript: https://github.com/SimHacker/moollm/blob/main/designs/LEELA-...
2. Architecture: Skills as Knowledge Units
A skill is a modular unit of knowledge that an LLM can load, understand, and apply.
Skills self-describe their capabilities, advertise when to use them, and compose with other skills.
Why Skills, Not Just MCP Tool Calls?
MCP (Model Context Protocol) tool calls are powerful, but each call requires a full round-trip:
MCP Tool Call Overhead (per call):
┌─────────────────────────────────────────────────────────┐
│ 1. Tokenize prompt │
│ 2. LLM complete → generates tool call │
│ 3. Stop generation, universe destroyed │
│ 4. Async wait for tool execution │
│ 5. Tool returns result │
│ 6. New LLM complete call with result │
│ 7. Detokenize response │
└─────────────────────────────────────────────────────────┘
× N calls = N round-trips = latency, cost, context churn
Skills operate differently. Once loaded into context, skills can:
Iterate:
MCP: One call per iteration
Skills: Loop within single context
Recurse:
MCP: Stack of tool calls
Skills: Recursive reasoning in-context
Compose:
MCP: Chain of separate calls
Skills: Compose within single generation
Parallel characters:
MCP: Separate sessions
Skills: Multiple characters in one call
Replicate:
MCP: N calls for N instances
Skills: Grid of instances in one pass
speed-of-light skill: https://github.com/SimHacker/moollm/tree/main/skills/speed-o...
Skills also compose. MOOLLM's cursor-mirror skill introspects Cursor's internals via a sister Python script that reads cursor's chat history and sqlite databases — tool calls, context assembly, thinking blocks, chat history. Everything, for all time, even after Cursor's chat has summarized and forgotten: it's still all there and searchable!
cursor-mirror skill: https://github.com/SimHacker/moollm/tree/main/skills/cursor-...
MOOLLM's skill-snitch skill composes with cursor-mirror for security monitoring of untrusted skills, also performance testing and optimization of trusted ones. Like Little Snitch watches your network, skill-snitch watches skill behavior — comparing declared tools and documentation against observed runtime behavior.
skill-snitch skil: https://github.com/SimHacker/moollm/tree/main/skills/skill-s...
You can even use skill-snitch like a virus scanner to review and monitor untrusted skills. I have more than 100 skills and had skill-snitch review each one including itself -- you can find them in the skill-snitch-report.md file of each skill in MOOLLM. Here is skill-snitch analyzing and reporting on itself, for example:
skill-snitch's skill-snitch-report.md: https://github.com/SimHacker/moollm/blob/main/skills/skill-s...
MOOLLM's thoughtful-commitment skill also composes with cursor-mirror to trace the reasoning behind git commits.
thoughtful-commit skill: https://github.com/SimHacker/moollm/tree/main/skills/thought...
MCP is still valuable for connecting to external systems. But for reasoning, simulation, and skills calling skills? In-context beats tool-call round-trips by orders of magnitude.
I'm not disagreeing with standards but instead of creating adapters, can't we prompt the agent to create its own version of a skill using its preferred guidelines? I don't think machines care about standards in the way that humans do. If we maintain pure knowledge in markdown, the agents can extract what they need on demand.
I see this as “libs” for the agents. They can discover relevant skills by searching a known index, extend their capability for on a given task.
While this makes sense for higher autonomy, it brings the well known supply chain security issue.
Currently all index list skills that are unverified. There are already examples of malicious skills.
I think skills are probably a net positive for the general population, but for power users, I do recommend moving one meta layer up --
Whenever there's an agent best practice (skill) or 'pre-prompt' that you want to use all the time, turn it into a text expansion snippet so that it works no matter where you are.
As an example, I have a design 'pre-prompt' that dictates a bunch of steering for agents re: how to pick style components, typography, layout, etc. It's a few paragraphs long and I always send it alongside requests for design implementation to get way-better-than-average output.
I could turn it into a skill, but then I'd have to make sure whatever I'm using supported skills -- and install it every time or in a way that was universally seen on my system (no, symlinking doesn't really solve this).
So I use AutoHotkey (you might use Raycast, Espanso, etc) to config that every time I type '/dsn', it auto-expands into my pre-prompt snippet.
Now, no matter whether I'm using an agent on the web/cloud, in my terminal window, or in an IDE, I've memorized my most important 'pre-prompts' and they're a few seconds away.
It's anti-fragile steering by design. Call it universal skill injection.
Please help me understand. Is a "skill" the prompt instructing the LLM how to do something? For example, I give it the "skill" of writing a fantasy story, by describing how the hero's journey works. Or I give it the "curl" skill by outputting curl's man page.
Its additional context that can be loaded by the agent as-needed. Generally it decides to load based on the skill's description, or you can tell it to load a specific skill if you want to.
So for your example, yes you might tell the agent "write a fantasy story" and you might have a "storytelling skill" that explains things like charater arcs, tropes, etc. You might have a separate "fiction writing" skill that defines writing styles, editing, consistency, etc.
All of this stuff is just 'prompt management' tooling though and isn't super commplicated. You could just paste the skill content into your context and go from there, this just provides a standardized spec for how to structure these on-demand context blocks.
Yes, pretty much.
LLM-powered agents are surprisingly human-like in their errors and misconceptions about less-than-ubiquitous or new tools. Skills are basically just small how-to files, sometimes combined with usage examples, helper scripts etc.
Started to work on a tool to synchronize all skills with symlinks. Its ok for my needs at the moment but feel free to improve it its on GH: https://github.com/Alpha-Coders/agent-loom
Thanks for the input I was planning to do something like that with a CLI though
I realized that amp uses ~/.agents/skills
I liked that idea to have something more CLI agnostic
Are there good techniques for testing / benchmarking skills effectiveness?
Awesome Agent Skills: https://github.com/skillmatic-ai/awesome-agent-skills
Comment was deleted :(
It's hilarious that after all those years of resistance to technical writing and formal specification engineers and programmers have suddenly been reduced to nothing more than technical writers and specification designers. Funny that I somehow don't foresee technical writing pay bumps happening as a consequence of this sudden surge in importance.
This post does a very good job of laying out that argument
Reads like slop to me. Reeeeeally verbose and many “it’s not just …, it’s …” all over the place.
slopman, the new strawman argument
You just don't know which parts of the doc are real and which are hallucinated. Maybe the prompter checked everything and the content is actually good, but sadly many don't and there is a lot of slop flying around.
So if you want others to read the output you'll have to de-slopify it, ideally make it shorter and remove the tells.
If I go by good faith these days and trust that someone didn't just upload llm hallucinated bullshit, I'd sadly just be reading slop all day and not learning anything or worse even get deceived by hallucinations and make wrong assumptions. It's just a waste of someones precious life time.
LLMs can read through slop all day, humans can not without getting extremely annoyed.
> You just don't know which parts of the doc are real and which are hallucinated.
It doesn't look like slop at all to me. GP claimed that this was written by ai without evidence, which I assumed to be based in bias, based on GP's comment history: https://news.ycombinator.com/threads?id=jondwillis They complaint they have about the writing style is not the style that is emblematic of Ai slop. Then, considering the depth of analysis and breadth of connection, this is not something current Ai is up to producing.
Are you also assuming the article was written by an Ai?
It's definitely LLM-written, or at least LLM-edited.
Uh, more like managers than writers. We (the agent and I) have written about 20 design docs for my personal project and none of them were by hand.
Is there a skill directory that can be browsed by a human?
I noticed a couple days ago https://skill.md started redirecting to this new URL.
Interesting format, but skills feel like optimizing the wrong layer. The agents usually don't fail because of bad instructions — they fail because external systems treat them like bots.
You can have the perfect scraping skill, but if the target blocks your requests, you're stuck. The hard problems are downstream.
If u wanna browse, search and download AI agent skills, use openskills.space
All these sites just look exactly like claude code skills doc.
Tank: We're supposed to start with these operation programs first. That's major boring shit. Let's do something a little more fun. How about... combat training.
Neo: Ju jitsu? I'm gonna learn Ju jitsu.
[Tank winks and loads the program] Neo: Holy shit!
KYS
Sounds like a bunch of bullshit to me. A simple markdown file with whatever and a directory will do the same. This is just packaging, selling and marketing.
Comment was deleted :(
[dead]
You've nailed the core insight about progressive disclosure. MOOLLM extends this into what we call the Semantic Image Pyramid (or MOO-Maps ;), borrowing from Mip-Maps in graphics. Four resolution levels, each serving different needs.
GLANCE.yml is the smallest, 5-70 lines. Just enough to answer "is this relevant?" You can inject all glances into every prompt because they're tiny. The LLM scans them like a table of contents.
CARD.yml is the interface layer, 50-200 lines. No implementation, just what the skill offers Capability advertisements, activation conditions, scoring, what it composes with. Think of it like The Sims "advertisement" system or CLOS generic function dispatch. The LLM sniffs this to decide whether to load the full SKILL.md implementation.
SKILL.md is the Anthropic-style skill file, 200-1000 lines. The actual instructions, the how. Only loaded when the skill is activated.
README.md is the largest, 500+ lines, and it's for humans. History, design rationale, examples. The LLM can dive in when developing the skill or when curious, but it's not burned on every invocation.
Reading rule: never load a lower level without first loading the level above. Start with GLANCE, sniff the CARD, load SKILL only if needed.
Even more compact than concatenating all glances:
We also found INDEX.md beats INDEX.yml for the skill catalog. YAML repeats the same keys for every entry. Markdown allows narrative explanation of how skills relate, which clusters matter for what tasks, making it both more compact and more useful.
INDEX.yml: 711 lines, 2061 words, 17509 chars, ~4380 tokens, machine readable structure
https://github.com/SimHacker/moollm/blob/main/skills/INDEX.y...
INDEX.md: 124 lines, 1134 words, 9487 chars, ~2370 tokens, human readable prose
https://github.com/SimHacker/moollm/blob/main/skills/INDEX.m...
INDEX.md is 83% fewer lines, 45% fewer words, 46% fewer chars for 121 skills. YAML repeats keys like id, tagline, why for every entry. Markdown uses headers and prose, compresses better, allows narrative grouping of related skills.
And it's simply more meaningful to both LLMs and humans, telling a coherent story instead of representing raw data!
The Semantic Image Pyramid:
https://github.com/SimHacker/moollm/blob/main/designs/LEELA-...
Same principle applies to code. A skill can wrap a sister-script that IS the documentation. A Python script with argparse defines the CLI once, readable by both humans (--help) and LLMs (sniff the top of the python file). No separate docs to maintain, no drift between what the code does and what the docs claim.
sister-script: https://github.com/SimHacker/moollm/blob/main/skills/sister-...
Sniffable-python structures code so the API is visible in the first 50 lines. Imports, constants, CLI structure up front. Implementation below the fold. The LLM can decide relevance and understand the API without reading the whole file. Single source of truth, progressive disclosure, don't repeat yourself.
sniffable-python README.md: https://github.com/SimHacker/moollm/blob/main/skills/sniffab...
sniffable-python SKILL.md: https://github.com/SimHacker/moollm/blob/main/skills/sniffab...
More: Speed of Light -vs- Carrier Pigeon (an allegory for Skills -vs- MCP):
https://github.com/SimHacker/moollm/blob/main/designs/SPEED-...
[dead]
[dead]
[dead]
[dead]
[dead]
[dead]
[dead]
[dead]
one good thing vercel did, was indexing skills.md under a site skills.sh - and yes there are now 100s of these sites, but I like the speedy/lite approach from vercel's DX, despite me not liking vercel a whole lot
I don't like vercel design, its just huge list of abstract skill name and you have to click on every one to even have a clue what something does. Such a bad design IMHO.
Design of https://www.skillcreator.ai/explore for me it's more useful. At least I can search by category, framework, language and I also see much more information what some skill does at a glance. I don't know why vercel really wanted to do it completely black and white - colors used and done with a taste gives useful context and information.
That site loads 1 skill at a time on the explore page on my iphone, mobile safari
slop?
Skill is an ability created through practice and habit. Text files have nothing to do with it - they're not skills.
I think pre-canned "skills" are an anti-pattern with the frontier models. Arguably, these skills already exist within the LLM. We don't need to explain how to do things they already know how to do.
I prefer to completely invert this problem and provoke the model into surfacing whatever desired behavior & capability by having the environment push back on it over time.
You get way more interesting behavior from agents when you allow them to probe their environment for a few turns and feed them errors about how their actions are inappropriate. It doesn't take very long for the model to "lock on" to the expected behavior if you are detailed in your tool feedback. I can get high quality outcomes using blank system prompts with good tool feedback.
I think skills actually complement what you're saying very well.
> You get way more interesting behavior from agents when you allow them to probe their environment for a few turns and feed them errors about how their actions are inappropriate. It doesn't take very long for the model to "lock on" to the expected behavior if you are detailed in your tool feedback. I can get high quality outcomes using blank system prompts with good tool feedback.
My primary way of developing skills (and previously cursor rules) is to start blank, let the LLM explore, and correct it as we go until the problem is solved. I then ask it to generate a skill (or rule) that explains the process in a way that it could refer to to repeat this again. Next time something like that comes up, we use the skill. If any correction is needed, I tell it to update the skill.
That way we get to have it explore and get more context initially, and then essentially "cache" that summarized context on the process for another time.
Error feedback from tools could be argued to be isomorphic with skills (or the development of them). It tracks with how we learn things in meatspace. Whatever strings we return in response to a bad SQL query or compiler error could also include the contents of some skill.md file.
What about libraries that are not in their training data? (e.g. new libraries, private libraries)
Or knowledge that is in their training data, but the majority of its training data isn't following the best practices? (e.g. Web Content Accessibility Guidelines)
I think there is a fair point in those cases of having a bunch of markdown docs files detailing them
Crafted by Rajat
Source Code