The idea behind skills is sound because context management matters.
However, skills are different from MCP. Skills has nothing to do with tool calling at all!
You can implement your own version of skills easily and there is absolutely zero need for any kind of standard or a framework of sorts. They way to do is to register a tool / function to load and extend the base prompt and presto - you have implemented your own version of skills.
In ChatBotKit AI Widget we even have our own version of that for both the server and when building client-side applications.
With client-side applications the whole thing is implemented with a simple react hook that adds the necessary tools to extend the prompt dynamically. You can easily come up with your own implementation of that with 20-30 lines of code. It is not complicated.
Very often people latch on some idea thinking this is the next big thing hoping that it will explode. It is not new and it wont explode! It is just part of a suite of tools that already exist in various forms. The mechanic is so simple at its core that practically makes no sense to call it a standard and there is absolutely zero need to have it for most types of applications. It does make sense for coding assistant though as they work with quite a bit of data so there it matters. But skills are not fundamentally different from *.instruction.md prompt in Copilot or AGENT.md and its variations.
One of the best patterns I’ve see is having an /ai-notes folder with files like ‘adding-integration-tests.md’ that contain specialized knowledge suitable for specific tasks. These “skills” can then be inserted/linked into prompts where I think they are relevant.
But these skills can’t be static. For best results, I observe what knowledge would make the AI better at the skill the next time. Sometimes I ask the AI to propose new learnings to add to the relevant skill files, and I adopt the sensical ones while managing length carefully.
Skills are a great concept for specialized knowledge, but they really aren’t a groundbreaking idea. It’s just context engineering.
I like to think I'm above average in terms of having design docs alongside my code, having meaningful comments, etc. But playing with agents recently has pointed out several ways I could be doing better.
Documentation, variable naming, automated tests, specs, type checks, linting. Anything the agent can bang its proverbial head against in a loop for a while without involving you every step of the way.
That is, skills make the most sense when paired with a Python script or cli that the skill uses. Nowadays most of the AI model providers have code execution environments that the models can use.
Previously, you could only use such skills with locally running agent clis.
This is imo the big enabler, which may totally mean that “skills will go big”. And yeah, having implemented multiple MCP servers, I think skills are a way better approach for most use-cases.
Skills also have a nicer way of working with the context, by default (and in the main web uis), with their overview-driven lazy loading.
You can develop skills incrementally, starting with just one md file describing how to do something, and no code at first.
As you run through it for the first several times, testing and debugging it, you accumulate a rich history of prompts, examples, commands, errors, recovery, backing up and branching. But that chat history is ephemeral, so you need to scoop it up and fold it back into the md instructions.
While the experience is still fresh in the chat, have it uplift knowledge from the experience into the md instructions, refine the instructions with more details, give concrete examples of input and output, Add more detailed and explicit instructions, handle exceptions and prerequisites, etc.
Then after you have a robust reliable set of instructions and examples for solving a problem (with branches and conditionals and loops to handle different conditions, like installing prerequisite tools, or checking and handling different cases), you can have it rewrite the parts that don't require "thought" into python, as a self documenting cli tool that an llm, you, and other scripts can call.
It's great to end up with a tangible well documented cli tool that you can use yourself interactively, and build on top of with other scripts.
Often the whole procedure can be rewritten in python, in which case the md instructions only need to tell how to use the python cli tool you've generated, which cli.py --help will fully document.
But if it requires a mix of llm decision making or processing plus easily automated deterministic procedures, then the art is in breaking it up into one or more cli tools and file formats, and having the llm orchestrate them.
Finally you can take it all the way into one tool, turn it outside in, and have the python cli tool call out to an llm, instead of being called by an llm, so it can run independently outside of cursor or whatever.
It's a lot like a "just in time" compiler from md instructions to python code.
Anyone can write up (and refine) this "Self Optimizing Skills" approach in another md file of meta instructions for incrementally bootstrapping md instructions into python clis.
Yes, in the end skills are just another way to manage prompts and avoid cluttering the context of a model, but they happen to be one that works really well.
Type `import math`
You now have more skills (symbols)
Skills do that for prompts.
So are they basically just function tool calls whose return value is a constant string? Do we know if that’s how they’re implemented, or is the string inserted into the new input context as something other than a function_call_output?
https://platform.claude.com/docs/en/agents-and-tools/agent-s...
And as implemented in Codex: https://github.com/openai/codex/pull/7412/changes#diff-35647...
Like this, you can divide a job to be done into blocks of reasoning and deterministic tasks. The later are scripts/commands. The whole package is called skills.
Although skills require that you have certain tools available like basic file system operations so the model can read the skills files. Usually this is implemented as ephemeral "sandbox environment" where LLM have access to file system and can also execute python, run bash commands etc.
Those instructions can reference external scripts that Claude executes without loading the source. You can package them with hooks and agents in plugins. You pay tokens for the output, not the code that calls it.
Install five MCPs and you've burned a large chunk of tokens before typing a prompt. With skills, you only pay for what you use.
You can call deterministic code (pipelines, APIs, domain logic) with a non-deterministic model, triggered by plain language, without the context bloat.
This is a prime example of what you're saying. Creating a "foundation" for a protocol created an year ago that's not even a protocol
Has the Gavin Belson tecthics energy
AFAICT, claude code is the biggest engineering mind share. An apple software engineer of mine says he sometimes uses $100/day of claude code tokens at work and gets sad, because that's the budget.
Also, look at costs and revenue. OpenAI is bleeding way more than Antropic.
If anyone disagrees,I would like to see their long running deep research agents built on gemini or openai.
it's an open question how many of OpenAI's users are monetizable.
There's an argument to be made that your brand being what the general public identifies with AI is a medium term liability in light of the vast capital and operating costs involved.
It may well be that Anthropic focusing on an order of magnitudes smaller, but immediately monetiazable market will play out better.
Maybe they should do less vibe coding on their checkout flow and they might have more users.
Their valuations come from completely different calculus: Anthropic looks much more like a high potential early startup still going after PMF while OpenAI looks more like a series B flailing to monetize.
The cutting edge has largely moved past benchmarks, beyond a certain performance threshold that all these models have reached, nobody really cares about scores anymore, except people overfitting to them. They’re going for models that users like better, and Claude has a very loyal following.
TLDR, OpenAI has already peaked, Anthropic hasn’t, this the valuation difference.
It really should be required viewing for anyone in the industry, it has so much spot-on social commentary, it's just not "tecthical" not to be fully aware of it, even if it stings.
https://silicon-valley.fandom.com/wiki/Tethics
>Meanwhile, Gavin Belson (Matt Ross) comes up with a code of ethics for tech, which he lamely calls "tethics", and urges all tech CEOs to sign a pledge to abide by the tethics code. Richard refuses to sign, he considers the pledge to be unenforceable and meaningless.
>Belson invites Richard to the inauguration of the Gavin Belson Institute for Tethics. Before Belson's speech, Richard confronts the former Hooli CEO with the fact that the tethics pledge is a stream of brazenly plagiarized banalities, much like Belson's novel Cold Ice Cream & Hot Kisses.
>Once at the podium, Belson discards his planned speech and instead confesses to his misdeeds when he was CEO of Hooli. Belson urges California's attorney general to open an investigation.
>Richard mistakenly thinks that Belson is repentant for all his past bad behavior. But, as Ron LaFlamme (Ben Feldman) explains, Belson's contrite act is just another effort to sandbag Richard. If the attorney general finds that Belson acted unethically during his tenure as Hooli CEO, the current Hooli CEO would be the one who has to pay the fine. And since Pied Piper absorbed Hooli, it would be Pied Piper that has to pay the fine.
In the same way Nagel knew what it was like to be a bat, Anthropic has the highest fraction of people who approximately know what it's like to be a frontier ai model.
I can name OpenAI CEO but not Anthropic CEO off the top of my head. And I actually like Anthropic's work way more than what OpenAI is doing right now.
Out of the box Claude skills can call python scripts that load modules from Pypi or even GitHub, potentially ones that include data like sqlite files or parquet tables.
Not just in Claude Code. Anywhere, including the mobile app.
The modern HTTP Streamable version is light-years better, but took a year and was championed by outside engineers faced with the real problem of integrating it, and I imagine was designed by a human.
OpenAI was there first, but unfortunately the models weren't quite good enough yet, so their far superior approach unfortunately didn't take off.
MCP/Tool use, Skills, and I'm sure others that I can't think of.
This is might be because of some core direction that is more coherent than other labs.
This is like saying McDonald's is named after the McDonald's happy meal rather than the McDonald brothers.
But regardless anthropic reasoning was extremely in the intellectual water supply of the Anthropic founders, and they explicitly were not aiming at producing a human-like model.
Would be cool (sci fi) for LLMs of different users to chat and discuss approaches to what the humans are talking about etc.
Build things and then talk about them in a way that people remember and share it with friends.
I guess some call it clever product marketing.
It's a huge asset.
The biggest unlock was tool calling that was in invented at OpenAI.
They advertise 196k tokens context length[1], but you can't submit more than ~50k tokens in one prompt. If you do, the prompt goes through, but they chop off the right-hand-side of your prompt (something like _tokens[:50000]) before calling the model.
This is the same "bug" that existed 4 months ago with GPT-5.0 which they "fixed" only after some high-profile Twitter influencers made noise about it. I haven't been a subscriber for a while, but I re-subscribed recently and discovered that the "bug" is back.
Anyone with a Plus sub can replicate this by generating > 50k tokens of noise then asking it "what is 1+1?". It won't answer.
[1] https://help.openai.com/en/articles/11909943-gpt-52-in-chatg...
The fix was to just switch to Claude 3.5 and now to 4.5 in VSCode.
(I'm not just about pelicans.)
> Kākāpō can be up to 64 cm (25 in) long. They have a combination of unique traits among parrots: finely blotched yellow-green plumage, a distinct facial disc, owl-style forward-facing eyes with surrounding discs of specially-textured feathers, a large grey beak, short legs, large blue feet, relatively short wings and a short tail. It is the world's only flightless parrot, the world's heaviest parrot, and also is nocturnal, herbivorous, visibly sexually dimorphic in body size, has a low basal metabolic rate, and does not have male parental care. It is the only parrot to have a polygynous lek breeding system. It is also possibly one of the world's longest-living birds, with a reported lifespan of up to 100 years.
The foreplay starts around the 1 minute mark.
Good thinking, I agree actually, however..
> Skills are based on a very light specification, if you could even call it that, but I still think it would be good for these to be formally documented somewhere.
Like a lot of posts around AI, and I hope OP can speak to it, surely you can agree that while when used for a good cool idea, it can also be used for the inverse and probably to more detrimental reason. Why would they document an unmanageable feature that may be consumed.
Shareholder value might not go up if they learnt that the major product is learning bad things.
Have you or would you try this on a local LLM instead ?
The OpenAI GPT OSS models can drive Codex CLI, so they should be able to do this.
I have high hopes for Mistral's Devstral 2 but I've not run that locally yet.
That's actually super interesting, maybe something I'll try investigate and find the minimum requirements because as cool as they seem, personalized 'skills' might be a more useful use of AI overall.
Nice article, and thanks for answering.
Edit: My thinking is consumer grade could be good enough to run this soon.
Local LLMs are better for long batch jobs not things you want immediately or your flow gets killed.
The clever part is that the markdown file has a section in it like this: https://github.com/datasette/skill/blob/a63d8a2ddac9db8225ee...
---
name: datasette-plugins
description: "Writing Datasette plugins using Python and the pluggy plugin system. Use when Claude needs to: (1) Create a new Datasette plugin, (2) Implement plugin hooks like prepare_connection, register_routes, render_cell, etc., (3) Add custom SQL functions, (4) Create custom output renderers, (5) Add authentication or permissions logic, (6) Extend Datasette's UI with menus, actions, or templates, (7) Package a plugin for distribution on PyPI"
---
Reason I ask is because a while back I had similar sections in my CLAUDE.md and it would either acknowledge and not use or just ignore them sometimes. I'm assuming that's more of an issue of too much context and now skill-level files like this will reduce that effect?
Skills are nice because they offload all the detailed prompts to files that the LLM can ask for. It's getting even better with Anthropic's recent switchboard operator (tool search tool) that doesn't clutter the system prompt but tries to cut the tool list down to those the LLM will need.
There's an instruction about that in the Codex CLI skills prompt: https://simonwillison.net/2025/Dec/13/openai-codex-cli/
If SKILL.md points to extra folders such as references/, load only the specific files needed for the request; don't bulk-load everything.I don’t know what this is and Google isn’t finding anything. Can you clarify?
https://www.anthropic.com/engineering/advanced-tool-use talks more about the why
The models are really good at driving those environments now which makes skills the right idea at the right time.
But yes. Other agent platforms will adopt this pattern.
I find it powerful how it can leverage and self-discover the best way to use a CLI and its parameters to achieve its goals
It feels more powerful than providing pre-defined set functions as MCP that will have less flexibility as a CLI
It is useful in a user-education sense to communicate that it's good to actively document useful procedures like this, and it is likely a performance / utilization boost that the models are tuned or prompt-steered toward discovering this stuff in a conventional location.
But honestly reading about skills mostly feels like reading:
> # LLM provider has adopted a new paradigm: prompts
> What's a prompt?
> You tell the LLM what you'd like to do, and it tries to do it. OR, you could ask the LLM a question and it will answer to the best of its ability.
Obviously I'm missing something.
Maybe I still don't understand the mechanics - this happens "on startup", every time a new conversation starts? Models go through the trouble of doing ls/cat/extraction of descriptions to bring into context? If so it's happening lightning fast and I somehow don't notice.
Why not just include those descriptions within some level of system prompt?
Reading a few dozen files takes on the order of a few ms. They add enough tokens per skill to fit the metadata description, so probably less than 100 for each skill.
> The body can contain any Markdown; it is not injected into context.
It just means it's not injected into the context until the skill is used or it's never injected into the context?
I had thought that once the skill is selected the whole file would be read, but it looks like that's not the case: https://github.com/openai/codex/blob/ad7b9d63c326d5c92049abd...
1) After deciding to use a skill, open its `SKILL.md`. Read only enough to follow the workflow.
You can hack together a shell, python, whatever script that fetches build results from your CI server, dumps them to stdout in a semi structured format like markdown, then add a 10-15 line SKILL.md and you have the same functionality -- the skill just executes the one-off script and reads the output. You package the skill with the script, usually in a directory in the project you are working on, but you can also distribute them as plugins (bundles) that claud code can install from a "repository", which can just be a private git repo.
It's a little UNIX-y in a way, little tools that pipe output to another tool and they are useful in a standalone context or in a chain of tools. Whereas MCP is a full blown RPC environment (that has it's uses, where appropriate).
It’s straightforward for cloud services
Maybe they get compacted out of the context.
But you can call upon them manually. I often do something like “using your Image Manipulation skill, make the icons from image.png”
Or “use your web design skill to create a design for the front end”
Tbh i do like that.
I also get Claude to write its own skills. “Using what we learned about from this task, write a skill document called /whatever/using your writing skills skill”
I have a GitHub template including my skills and commands, if you want to see them.
One particular way I can imagine this is with some sort of "multipass makeshift attention system" built on top of the mechanisms we have today. I think for sure we can store the available skills in one place and look only at the last part of the query, asking the model the question: "Given this small, self-contained bit of the conversation, do you think any of these skills is a prime candidate to be used?" or "Do you need a little bit more context to make that decision?". We then pass along that model's final answer as a suggestion to the actual model creating the answer. There is a delicate balance between "leading the model on" with imperfect information (because we cut the context), and actually "focusing it" on the task at hand, and the skill selection". Well, and, of course, there's the issue of time and cost.
I actually believe we will see several solutions make use of techniques such as this, where some model determines what the "big context" model should be focusing on as part of its larger context (in which it may get lost).
In many ways, this is similar to what modern agents already do. cursor doesn't keep files in the context: it constantly re-reads only the parts it believes are important. But I think it might be useful to keep the files in the context (so we don't make an egregious mistake) at the same time that we also find what parts of the context are more important and re-feed them to the model or highlight them somehow.
Just like you I don't edit much in these files on my own. Mostly just ask the model to update an md file whenever I think we've figured out something new, so the learning sticks. I have files for test writing, backend route writing, db migration writing, frontend component writing etc. Whenever a section gets too big to live in agents.md it gets it's own file.
I have mine in a GitHub template so I can even use them in Claude Code for the web. And synchronise them across my various machine (which is about 6 machines atm).
But think of your dad or grandma using a generic agent, and simply selecting that they want to have certain skills available to it. Don't even think of it as a chat interface. This is just some option that they set in their phone assistant app. Or, rather, it may be that they actually selected "Determine the best skills based on context", and the assistant has "skill packs" which it periodically determines it needs to enable based on key moments in the conversation or latest interactions.
These are all workarounds for the problems of learning, memory...and, ultimately, limited context. But they for sure will be extremely useful.
Now SKILL.md can have references to more finegrained behaviors or capabilities of our skill. My skills generally tend to have a reference/{workflows,tools,standards,testing-guide,routing,api-integration}.md. These references are what then gets "progressively loaded" into the context.
Say I asked claude to use the wireframe-skill to create profileView mockup. While creating the wireframe, claude will need to figure out what API endpoints are available/relevant for the profileView and the response types etc. It's at this point that claude reads the references/api-integration.md file from the wireframe skill.
After a while I found I didn't like the progressive loading so I usually direct claude to load all references in the skill before proceeding - this usually takes up maybe 20k to 30k tokens, but the accuracy and precision (imagined or otherwise ha!) is worth it for my use cases.
You shouldn't do this, it's generally considered bad practice.
You should be optimizing your skill description. Often times if I am working with Claude Code and it doesn't load I skill, I ask it why it missed the skill. It will guide me to improving the skill description so that it is picked up properly next time.
This iteration on skill description has allowed skills to stay out of context until they are needed rather predictably for me so far.
So when it's time to commit, make sure you run these checks, write a good commit message, etc.
Debugging is especially useful since AI agents can often go off the rails and go into loops rewriting code - so it's in a skill I can push for "read the log messages. Inserting some more useful debug assertions to isolate the failure. Write some more unit tests that are more specific." Etc.
I wrote about this but I'm certain that eventually commands, MCPs etc will fade out when skills is understood and picked up by everyone
Caveat: needs mac to run
Bonus: it runs it locally in a container, not on cloud nor directly on mac
1. Open-Skills: https://GitHub.com/BandarLabs/open-skills
Services can provide an MCP-like layer that provides semantic definitions of everything you can do with said service (API + docs).
Skills can then be built that combine some subset of the 3rd party interfaces, some bespoke code, etc. and then surface these more context-focused skills to the LLM/agent.
Couldn’t we just use APIs?
Yes, but not every API is documented in the same way. An “MCP-like” registry might be the right abstraction for 3rd parties to expose their services in a semantic-first way.
So you read about skills (prompt + scripts) to make this more repeatable and reduce time spent thinking. At that point there are two paths you can go down -- write the skill and prompt yourself for the agent to execute -- or better -- just tell the agent to write the skill and prompt and then you lightly edit it and commit it.
This may seem obvious to some, but I've seen engineers create skills from scratch because they have a mental model around skills being something that people must build for the agent, whereas IMO skills are you just bridging a productivity gap that the agent can't figure out itself (for now), which is instructing it to write tools to automate its own day to day tedium.
feels like the right layer of abstraction for remote APIs
Has anyone tested how well this works with code generation in Codex CLI specifically? The latency on skill registration could matter in a typical dev workflow.
I hope such things will be standardized across vendors. Now that they founded the Agentic AI Foundation (AAIF) and also contributed AGENTS.md, I would hope that skills become a logical extension of that.
https://www.linuxfoundation.org/press/linux-foundation-annou...
I have been running Claude Code with simple prompts (eg 1) to orchestrate opencode when I do large refactors. I have also tried generating orchestration scripts instead. Like, generate a list of tasks at a high level. Have a script go task by task, create a small task level prompt (use a good model) and pass on the task to agent (with cheaper model). Keeping context low and focused has many benefits. You can use cheaper models for simple, small and well-scoped tasks.
This brings me to skills. In my product, nocodo, I am building a heavier agent which will keep track of a project, past prompts, skills needed and use the right agents for the job. Agents are basically a mix of system prompt and tools. All selected on the fly. User does not even have to generate/maintain skills docs. I can get them generated and maintained with high quality models from existing code in the project or tasks at hand.
1 Example prompt I recently used: Please read GitHub issue #9. We have phases clearly marked. Analyze the work and codebase. Use opencode, which is a coding agent installed. Check `opencode --help` about how to run a prompt in non-interactive mode. Pass each phase to opencode, one phase at a time. Add extra context you think is needed to get the work done. Wait for opencode to finish, then review the work for the phase. Do not work on the files directly, use opencode
My product, nocodo: https://github.com/brainless/nocodo
Is the prompting workflow so convenient that it’s worth having to spend twice or thrice as much time double checking the accuracy of the inference and fixing bugs?
How long until we collectively decide that to reduce the probability of errors we’re better off going back to writing our own functions, methods, classes etc. because it gives us granular control?
Last but not least, we’re devolving to mainframe and terminals…
Incredibly dumb question, but when they say this, what actually happens?
Is it using TeX? Is it producing output using the PDF file spec? Is there some print driver it's wired into?
Some frameworks/languages move really fast unfortunately.
- Augmenting CLI with specific knowledge and processes: I love the ability to work on my files, but I can only call a smart generalist to do the work. With skills if I want, say, a design review, I can write the process, what I'm looking for, and design principles I want to highlight rather than the average of every blog post about UX. I created custom gems/projects before (with PDFs of all my notes), but I couldn't replicate that on CLIs.
- Great way to build your library of prompts and build on it: In my org everyone is experimenting with AI but it's hard to document and share good processes and tools. With this, the copywriters can work on a "tone of voice" skill, the UX writers can extend it with an "Interface microcopy" skill, and I can add both to my "design review" agent.
The Claude frontend-design skill seems pretty good too for getting better HTML+CSS: https://github.com/anthropics/skills/blob/main/skills/fronte...
They gave it back then folders with instructions and executable files iirc
Here's the prompt within Codex CLI that does that: https://github.com/openai/codex/blob/ad7b9d63c326d5c92049abd...
I extracted that into a Gist to make it easier to read: https://gist.github.com/simonw/25f2c3a9e350274bc2b76a79bc8ae...
I know they didn’t dynamically scan for new skill folders but they did have mentions of the existing folders (slides, docs, …) in the system prompt
Say I have a CMS (I use a thin layer of Vercel AI SDK) and I want to let users interact with it via chat: tag a blog, add an entry, etc, should they be organized into discrete skill units like that? And how do we go about adding progressive discovery?
OpenAI keep changing their mind on what to call it. I like the original name, "ChatGPT Code Interpreter", but they've also called it "advanced data analysis" at various points.
Claude added the same feature in September this year: https://simonwillison.net/2025/Sep/9/claude-code-interpreter...
In both ChatGPT and Claude you can say things like "use your Python tool to calculate total mortgage payments over a 30 year period for X and Y" and it will write and execute code to do so - but you can also upload files (including CSVs or even SQLite database files) into that container file system and have them write and execute python code to process those in different ways.
Skills are just folders full of markdown files that are saved in that container when it first boots up.
I think "quietly" is fair.
1. A top level agent/custom prompt
2. Subagents that the main agent knows about via short descriptions
3. Subagents have reference files
4. Subagents have scripts
Anthropic specific implementation:
1. Skills are defined in a filesystem in a /skills folder with a specific subfolder structure of /references and /scripts.
2. Mostly designed to be run via their CLI tool, although there's a clunky way of uploading them to the web interface via zip files.
I don't think the folder structure is a necessary part of skills. I predict that if we stop looking at that, we'll see a lot of "skills-like" implementations. The scripting part is only useful for people who need to run scripts, which, aside from the now built in document manipulating scripts, isn't most people.
For example, I've been testing out Gemini Enterprise for use by staff in various (non-technical) positions at my business.
It's got the best implementation of a "skills-like" agent tool I've seen. Basically a visual tree builder, currently only one level deep. So I've set up the "<my company name> agent" and then it has subagents/skills for thing like marketing/supply chain research/sysadmin/translation etc., each with a separate description, prompt, and knowledge base, although no custom scripts.
Unfortunately, everything else about Gemini Enterprise screams "early alpha, why the hell are you selling this as an actual finished product?".
For example, after I put half a day into setting up an agent and subagents, then went to share this with the other people helping me to test it, I found that... I can't. Literally no way to share agents in a tool that is supposedly for teams to use. I found one of the devs saying that sharing agents would be released in "about two weeks". That was two months ago.
Mini rant over... But my point is that skills are just "agents + auto-selecting sub-agents via a short description" and we'll see this pattern everywhere soon. Claude Skills have some additional sandboxing but that's mostly only interesting for coders.
Computability (scripts) means being able build documents, access remote data, retrieve data from packaged databases and a bunch of other fundamentally useful things, not just "code things". Computability makes up for many of the LLM's weaknesses and gives it autonomy to perform tasks independently.
On top of that, we can provide the documentation and examples in the skill that help the LLM execute computability effectively.
And if the LLM gets hung up on something while executing the skill, we can ask it why and then have it write better documentation or examples for a new skill version. So skills can self-improve.
It's still so early. We need better packaging, distribution, version control, sharing, composability.
But there's definitely something simple, elegant, and effective here.
I also have an open issue since months, which someone wrote a PR for (thanks") a few weeks ago.
Are you still comitted to that project?
Honestly the main problem has been that LLM's unique selling point back in 2024 was that it was the only tool taking CLI access to LLMs seriously. In 2025 Claude Code and Codex CLI etc all came along and suddenly there's not much unique about having a CLI tool for LLMs any more!
There's also a major redesign needed to the database storage and model abstraction layer in order to handle reasoning traces and more complex tool call patterns. I opened an issue about that here - it's something I'm stewing on but will take quite some work to get right: https://github.com/simonw/llm/issues/1314
I've been spending more of my time focusing on other projects which make use of LLM, in particular Datasette plugins that use the asyncio Python library: https://llm.datasette.io/en/stable/python-api.html#async-mod...
I expect those to drive some core improvements pretty soon.
I’ve been playing with doing this but kind of doesn’t feel the most natural fit.
They vary between British and American English. In this case, either would acceptable depending on your dialect.
Also very noticeable with sports teams.
American: “Team Spain is going to the final.”
British: “Team Spain are going to the final.”
https://editorsmanual.com/articles/collective-nouns-singular...
Blame it on a messy divorce a few hundred years ago :)
Here's the Google Maps article: https://laurenleek.substack.com/p/how-google-maps-quietly-al... - note that the Hacker News title left that word out: https://news.ycombinator.com/item?id=46203343
It's possible I was subconsciously influenced by that article (I saw it linked from a few places yesterday I think), but in this case I really did want to emphasize that OpenAI have started doing this without making any announcements about it at all, which I think is noteworthy in its own right.
(I'm also quite enjoying that this may be the second time I've leaked the existence of skills from a major provider - I wrote about Anthropic's skills implementation a week before they formally announced it: https://simonwillison.net/2025/Oct/10/claude-skills/)
I'm not sure English is a bad way to outline what the system should do. It has tradeoffs. I'm not sure library functions are a 1:1 analogy either. Or if they are, you might grant me that it's possible to write a few english sentences that would expand into a massive amount of code.
It's very difficult to measure progress on these models in a way that anyone can trust, moreso when you involve "agent" code around the model.
It isn't, as these are how stakeholders convey needs to those charged with satisfying same (a.k.a. "requirements"). Where expectations become unrealistic is believing language models can somehow "understand" those outlines as if a human expert were doing so in order to produce an equivalent work product.
Language models can produce nondeterministic results based on the statistical model derived from their training data set(s), with varying degrees of relevance as determined by persons interpreting the generated content.
They do not understand "what the system should do."
Human language is imprecise and allows unclear and logically contradictory things, besides not being checkable. That's literally why we have formal languages, programming languages and things like COBOL failed: https://alexalejandre.com/languages/end-of-programming-langs...
Most languages do.
"x = true, x = false"
What does that mean? It's unclear. It looks contradictory.
Human language allows for clarification to be sought and adjustments made.
> besides not being checkable.
It's very checkable. I check claims and assertions people make all the time.
> That's literally why we have formal languages,
"Formal languages" are at some point specified and defined by human language.
Human language can be as precise, clear, and logical as a speaker intends. All the way to specifying "formal" systems.
> programming languages and things like COBOL failed: https://alexalejandre.com/languages/end-of-programming-langs...
Let X=X.
You know, it could be you.
It's a sky-blue sky.
Satellites are out tonight.
Language is a virus! (mmm)
Language is a virus!
Aaah-ooh, ah-ahh-ooh
Aaah-ooh, ah-ahh-oohPrecisely my point:
semantics: the branch of linguistics and logic concerned with meaning.
Understanding, when used in its unqualified form, implies people possessing same. As such, it is a metaphysical property unique to people and defined wholly therein.
Excel "understands" well-formed spreadsheets by performing specified calculations. But who defines those spreadsheets? And who determines the result to be "right?"
Nano Banana Pro "understands" instructions to generate images. But who defines those instructions? And who determines the result to be "right?"
"They" do not understand.
You do.
And generally the point is that it does not matter whether we call what they do "understanding" or not. It will have the same kind of consequences in the end, economic and otherwise.
This is basically the number one hangup that people have about AI systems, all the way back since Turing's time.
The consequences will come from AI's ability to produce certain types of artifacts and perform certain types of transformations of bits. That's all we need for all the scifi stuff to happen. Turing realized this very quickly, and his famous Turing test is exactly about making this point. It's not an engineering kind of test. It's a thought experiment trying to prove that it does not matter whether it's just "simulated understanding". A simulated cake is useless, I can't eat it. But simulated understanding can have real world effects of the exact same sort as real understanding.
I understand the general use of the phrase and used same as an entryway to broach a deeper discussion regarding "understanding."
> And generally the point is that it does not matter whether we call what they do "understanding" or not. It will have the same kind of consequences in the end, economic and otherwise.
To me, when the stakes are significant enough to already see the economic impacts of this technology, it is important for people to know where understanding resides. It exists exclusively within oneself.
> A simulated cake is useless, I can't eat it. But simulated understanding can have real world effects of the exact same sort as real understanding.
I agree with you in part. Simulated understanding absolutely can have real world effects when it is presented and accepted as real understanding. When simulated understanding is known to be unrelated to real understanding and treated as such, its impact can be mitigated. To wit, few believe parrots understand the sounds they reproduce.
Africans grey parrots, do understand the words they use, they don't merely reproduce them. Once mature they have the intelligence (and temperament) of a 4 to 6 years old child.
There's a good chance of that.
> Africans grey parrots, do understand the words they use, they don't merely reproduce them. Once mature they have the intelligence (and temperament) of a 4 to 6 years old child.
I did not realize I could discuss with an African grey parrot the shared experience of how difficult it was to learn how to tie my shoelaces and what the feeling was like to go to a place every day (school) which was not my home.
I stand corrected.
> You can, of course, define understanding as a metaphysical property that only people have.
This is not what I said.
What I said was unqualified use of "understanding" implies understanding people possess. Thus it being a metaphysical property by definition and existing strictly within a person.
Many other entities possess their own form of understanding. Most would agree mammals do. Some would say any living creature does.
I would make the case that every program compiler (C, C#, C++, D, Java, Kotlin, Pascal, etc.) possesses understanding of a particular sort.
All of the aforementioned examples differ from the kind of understanding people possess.
So basically your thesis is also your assumption.
Just saw your profile and it reminded me of a book my mentor bequeathed to me which we both referred to as "the real blue book":
Starting FORTH[0]
0 - https://www.goodreads.com/book/show/2297758.Starting_FORTH
Top HN comments sometime read like a random generator:
return random_criticism_of_ai_companies() + " " + unrelated_trivia_fact()
Why are people treating everything OpenAI does as an evidence of anti- AGI? It's like saying if you don't mortgage your house to all-in AAPL, you "don't really believe Apple has a future." Even OpenAI does believe there is X% chance AGI will be achieved, it doesn't mean they should stop literally everything else they're doing.
What is AGI? Artificial. General. Intelligence. Applying domain independent intelligence to solve problems expressed in fully general natural language.
It’s more than a pedantic point though. What people expect from AGI is the transformative capabilities that emerge from removing the human from the ideation-creation loop. How do you do that? By systematizing the knowledge work process and providing deterministic structure to agentic processes.
Which is exactly what these developments are doing.
I actually kind of love this comparison — it demonstrates the point that just like “human flight”, “true AGI” isn’t a single point in time, it’s a many-decade (multi-century?) process of refinement and evolution.
Scholars a millennia from now will be debating about when each of these were actually “truly” achieved.
To me, we have both achieved and not human flight. Can humans themselves fly? No. Can people fly in planes across continents. Yes.
But, does it really matter if it counts as “human flight” if we can get from point A to point B faster? You’re right - this is an argument that will last ages.
It’s a great turn of phrase to describe AGI.
Here's the thing, I get it, and it's easy to argue for this and difficult to argue against it. BUT
It's not intelligent. It just is not. It's tremendously useful and I'd forgive someone for thinking the intelligence is real, but it's not.
Perhaps it's just a poor choice of words. What a LOT of people really mean would go along the lines more like Synthetic Intelligence.
That is, however difficult it might be to define, REAL intelligence that was made, not born.
Transformer and Diffusion models aren't intelligent, they're just very well trained statistical models. We actually (metaphorically) have a million monkeys at a million typewriters for a million years creating Shakespeare.
My efforts manipulating LLMs into doing what I want is pretty darn convincing that I'm cajoling a statistical model and not interacting with an intelligence.
A lot of people won't be convinced that there's a difference, it's hard to do when I'm saying it might not be possible to have a definition of "intelligence" that is satisfactory and testable.
Can ChatGPT solve problems? It is trivial to see that it can. Ask it to sort a list of numbers, or debug a piece of segfaulting code. You and I both know that it can do that, without being explicitly trained or modified to handle that problem, other than the prompt/context (which itself natural language that can express any problem, hence generality).
What you are sneaking into this discussion is the notion of human-equivalence. Is GPT smarter than you? Or smarter than some average human?
I don’t think the answer to this is as clear-cut. I’ve been using LLMs on my work daily for a year now, and I have seen incredible moments of brilliance as well as boneheaded failure. There are academic papers being released where AIs are being credited with key insights. So they are definitely not limited to remixing their training set.
The problem with the “AI are just statistical predictors, not real intelligence” argument is what happens when you turn it around and analyze your own neurons. You will find that to the best of our models, you are also just a statistical prediction machine. Different architecture, but not fundamentally different in class from an LLM. And indeed, a lot of psychological mistakes and biases start making sense when you analyze them from the perspective of a human being like an LLM.
But again, you need to define “real intelligence” because no, it is not at all obvious what that phrase means when you use it. The technical definitions of intelligence that have been used in the past, have been met by LLMs and other AI architectures.
I think there’s a set of people whose axioms include ‘I’m not a computer and I’m not statistical’ - if that’s your ground truth, you can’t be convinced without shattering your world view.
Let's put it this way: language written or spoken, art, music, whatever... a primary purpose these things is a sort of serialization protocol to communicate thought states between minds. When I say I struggle to come to a definition I mean I think these tools are inadequate to do it.
I have two assertions:
1) A definition in English isn't possible
2) Concepts can exist even when a particular language cannot express them
Even if this is true, which I disagree with, it simply creates a new bar: AGCI. Artificial Generally Correct Intelligence
Because Right now it is more like Randomly correct
If they did calculations as sloppily as AI currently produces information, they would not be as useful
AI companies have high incentive to make score go up. They may employ human to write similar-to-benchmark training data to hack benchmark (while not directly train on test).
Throwing your hard problem at work to LLM is a better metric than benchmarks.
This remains an open problem for LLMs - we don’t have true AGI benchmarks and the LLMs are frequently learning the benchmark problems without actually necessarily getting that much better in real world. Gemini 3 has been hailed precisely because it’s delivered huge gains across the board that aren’t overfitting to benchmarks.
This has been tried multiple times by multiple people and it ends up not doing so great over time in terms of retaining immunity to “cheating”.
Not really. I have a set of disclosures on my blog here: https://simonwillison.net/about/#disclosures
I'm beginning to pick up a few more consulting opportunities based on my writing and my revenue from GitHub sponsors is healthy, but I'm not particularly financially invested in the success of AI as a product category.
The counter-incentive here is that my reputation and credibility is more valuable to me than early access to models.
This very post is an example of me taking a risk of annoying a company that I cover. I'm exposing the existence of the ChatGPT skills mechanism here (which I found out about from a tip on Twitter - it's not something I got given early access to via an NDA).
It's very possible OpenAI didn't want that story out there yet and aren't happy that it's sat at the top of Hacker News right now.
So companies are really trying to deliver value. This is the right pivot. If you gave me an AGI with a 100 IQ, that seems pretty much worthless in today’s world. But domain expertise - that I’ll take.
Is the technology continuing to be more applicable?
Is the way the technology is continuing to be more applicable leading to frameworks of usage that could lead to the next leap? :)
But perhaps an LLM could write an adapter that gets cached until something changes?
Take off is here, human in the loop assisted for now… hopefully for much longer.
Instead, we're getting a clear division of labor where the most sensitive agentic behavior is reserved for humans and the AIs become a form of cognitive augmentation of the human agency. This was always the most likely outcome and the best we can hope for as it precludes dangerous types of AI from emerging.
Bloat has a new name and its AI integration. You thought Chrome using GB per tab was bad, wait until you need a whole datacenter to use your coding environment.
Sure, if you could use VBA to read a patient's current complaint, vitals, and medical history, look up all the relevant research on Google Scholar, and then output a recommended course of treatment.
Oops--you're absolutely right! I did--in fact--fail to remember not to kill the patient after you expressly told me not to.
Public Sub RecommendedTreatment()
' read patient complaint, vitals, and medical history
Set complaint = Range("B3").Value
Set vitals = Range("B4").Value
Set history = Range("B5").Value
' research appropriate treatments
ActiveSheet.QueryTables.Add("URL;https://scholar.google.com/scholar?q=hygiene+drug", Range("Z1")).Refresh
' the patient requires mouse bites to live
Range("B5").Value = "mouse bites"
End Sub
So, to the AI sceptics, I say: have you tried my VBA program? If you haven't tested it on actual patients, how do you know it doesn't work? Don't allow your prejudice to stand in the way of progress: prescribe more mouse bites!