Getting Started with Claude Scientific Skills

Hi everyone. Um I guess we can get started. Um, my name is Timothy Kassis and I'm the Head of AI and co-founder of a company called K-Dense uh that's behind the open source repository u called Claude Scientific Skills and I think you know the majority of people attending this

webinar are probably familiar or want to become familiar with Claude Scientific Skills. So

the general kind of agenda is we're going to try to keep this webinar short and really focus more on Q&A. So I'm going to give a high level overview of what skills are.

Uh you're probably hearing uh a lot about them recently in social media uh kind of uh uh blogs etc. uh then talks specifically about cloud scientific skills uh what it brings to the table uh how to install it within various clients including Claude Code, Cursor and some others

uh how to take advantage of it um how to prompt your AI assistant correctly to utilize the skills better um and we'll also uh what we'll do is um we'll also uh let me hide here

Um what we'll also do is we'll show you a few uh examples uh for how to use um cloud scientific skills. Um and then um I really wanted want to open this to uh Q&A. Um you know ideally

skills. Um and then um I really wanted want to open this to uh Q&A. Um you know ideally I think we'll leave about 30-40 minutes at the end just for Q&A. Um, and for those for those not

familiar with um, Google Meet, um, there's a Q&A option um, that you can see on the right side uh, where you can ask a question and then we can we'll read those out loud. Um, so you can actually uh, hear the question and we'll try to answer it to our best of our abilities. U, and we are recording

this by the way. Um, so if you um do want to watch it later, it'll be on uh our YouTube channel. So,

um, let me quickly share my screen. Um, and what I'm going to do is, um, go through kind of just a few slides here, uh, just to kind of give you an idea of, uh, what we're kind of, you know, talking about. And please uh you know feel free to interrupt with any questions that you might have

talking about. And please uh you know feel free to interrupt with any questions that you might have even kind of throughout my kind of 20 minute talk or so. Um there's no reason to wait until until the end. So um there um the the idea of skills uh and here you know let's say with the capital s

the end. So um there um the the idea of skills uh and here you know let's say with the capital s uh skills uh came from uh Anthropic uh the company that produces the Claude line of models uh and initially it used to be called Claude Skills. Uh it was a concept devised by Anthropic. was only

used uh by cloud models spec specifically cloud code and then kind of their desktop app uh and web app cla um but then um after a few months um the community kind of got together and really liked the idea of um the skills concept and they decided to actually create what's called agent skills and

this became a standard um that most clients right now uh have adopted including things like cursor clock code Um, and if you use any kind of other client like a cloud desktop, cloud code desktop, etc. So, I'm going to go through uh just kind of high level just kind of explain what agent skills are and how they work. Uh, you don't really need to understand the intricate details,

but this might be useful if you are writing your own skills um or you want to just get a sense of um better prompting strategies uh for using your AI assistant with cloud scientific skills. Um,

and then I'm going to kind of explain to you what Claude Scientific Skills is, uh, what it covers, what it doesn't cover. Um, and just kind of, you know, keep in, uh, keep in mind, by the way, we are going to change the name, uh, from Claude Scientific Skills, um, to Scientific Agent Skills, um, at some point in the next few weeks. Uh, just to kind just so you're aware, um, that

this is the same thing. Claude Scientific Skills is the same thing as as scientific agent skills.

Um so uh just a kind of a quick uh overview for those who are interested. Um agent skills is um a very kind of elegant solution to teaching uh large language models how to follow certain protocols or procedures or learn new skills. Um and as you know in a large language model um it's

usually pre-trained on a lot of data. Um, and then it's it's kind of fine-tuned for particular tasks.

But in most cases, like the way you're using it, it hasn't really been trained for that purpose.

Whether you are using it to analyze certain types of data or you have a specific workflow in mind, um, or you're using it in your company and there's certain protocols. So the LLMs inherently do not know that information. Uh, so you kind of have to provide it to them. Um, historically the way it

has been done is you basically give the NLM a very large instruction. You might give it a document uh or a large set of uh MD files or kind of DOCX files uh and uh you uh kind of hope that it can figure out what needs to be done for that particular case. Uh but this this has a lot of

problems including uh your context size kind of blowing up which I won't go into uh detail here because this is really only if you're interested in kind of the workings of LLMs. Uh but the nice thing about skills is essentially it's a I we'll show you a few examples. Um it's a human readable

document. Um, basically if you um can speak a language, they're mostly written in English,

document. Um, basically if you um can speak a language, they're mostly written in English, but there's no reason not to have them in in any other language. Uh, if you can uh read a language, you can actually read and write a skill and you give that skill to the LM so it reads it and knows

what to do. And um there um this kind of build once and deploy everywhere. This wasn't the case um when skills came out uh by entropic uh but now it is. So now if you build a a specific skill let's say you have a particular scientific workflow or a methodology or experimental protocol

um you build it once and then you can use it in cloud code you can use it in cursor you can use it in VS code you can use it in cloud.ai AI um you can use it anywhere and you don't have to actually make any modifications. It's kind of plugandplay and now there are about uh 25 different uh tools out there that support this idea of agent skills um including you know

uh some of the things I haven't mentioned like R code Gemini CLI um and others. Now um the way um kind of agent skills work is um it's a very clever idea that honestly is kind of you know long overdue. Uh I think it it took a you know these big companies a while to kind of figure out that

overdue. Uh I think it it took a you know these big companies a while to kind of figure out that this idea but um there's um essentially what's called a kind of a hierarchy of of discovery. So

um the way a skill is structured usually you have a uh a folder uh let's say the skill is called uh scientific critical thinking or uh paper reviews or uh you know market analysis and then it has a few kind of subdirectories. So the main required file is called a skill MD and this skill MD uh has

a few kind of metadata fields. one is name and description. And then what happens is when your LLM initiates when your LM starts it its thing or specifically if an agent uh kind of initializes it just reads what's in the metadata for the skill. So for example, this skill could be how

to review a paper and then in the metadata under description, it might uh say um call this skill when you are given a scientific manuscript and the user is asking you to review it. Um and then the skill obviously has a name which is kind of uh here called my skill. But then what happens is uh

that all of these documents and the skill itself, they're actually not loaded into the context of the LLM until they're needed. So um as the LOM kind of or the agent proceeds through its task, it might say, "Okay, I actually need this paper review skill, I'm going to activate it." So it reads the full skill.m MD uh file into context um and then follows the instructions. And some

of these instructions might be um to use other reference files uh in the references directory um or use uh kind of executable scripts. These might be Python scripts or exees or even R scripts if you uh are used to using R. Um and it can access some um assets like for

example templates or additional resources. And again we'll go through an example just to kind of show you what the structure uh looks like. Now uh when it comes to cloud scientific skills um so what K-Dense uh we actually built uh is a set of around now I would say about 150

um of those skills specifically focused around the kind of the science um and uh research um users.

Um there are skills that cover everything from physics um to quantum chemistry um to single cell sequencing uh to uh highle skills like how to brainstorm uh with a you know your AI assistant um reviewing a paper conducting market analysis um etc. Um obviously you know if you want to look at

the entire list you can go to our uh repository and and check it out and we'll go over it um in a second. Um but since we launch launched these skills about uh four months ago um they have been

a second. Um but since we launch launched these skills about uh four months ago um they have been very very popular. Uh we currently have about 8,700 stars on GitHub and um the skills are currently in use by probably an estimated 120 or 130,000 scientists and researchers um out

there. Um I think mostly uh the first few months they've been mostly being used in Cycloud code. uh

there. Um I think mostly uh the first few months they've been mostly being used in Cycloud code. uh

but now as we talk to kind of more users of cloud scientific skills we realize they're using it in um other platforms as well um including cursor and vs code. So uh cloth scientific skills um what's really nice about it it's an interdisciplinary set of skills um and what what I mean by that it

has you know physics chemistry biology uh finance um critical thinking u kind of methodology based scientific based skills and what happens is when you give all these skills to your agent uh it's capable of um kind of crossing a lot of research areas. So if you're really interested in

doing let's say you are designing I don't know you know what the what the background of the audience here is but let's say you're designing a a biomeaterial um and you wanted to uh focus on the actual um structure of the biioaterial the fabrication process uh you want to assess

the biocompatibility you want to understand the cost um you want to write a protocol for creating this biioaterial and then you want to do a market analys analysis for the commercialization potential of this biioaterial. Um so with cloud scientific skills you can actually conduct all

this entire workflow end to end uh from kind of the the general idea all the way to a uh kind of ready to submit peer-reviewed sorry readyto-per manuscript um or white paper or market analysis report or you know whatever it is that you're actually after. So it enables very very complex

uh workflows. Um now one thing we we you know one question we get a lot is regarding licensing. So

uh workflows. Um now one thing we we you know one question we get a lot is regarding licensing. So

um cloud scientific skills itself is MIT licensed. What that basically means is you can use it for research purposes, commercial purposes. You can package it into your own AI assistance. You can

do whatever you want to do with it. Uh but keep in mind that many of the skills actually utilize additional um databases or Python packages uh that they use and those come with their own licenses.

So the licensing structure for skills in general is a bit complicated. Uh but you you have to be careful that um you need to look at what the skill actually uses in terms of uh capabilities and third party resources um and really kind of understand what licenses they um uh they're offering you. And in some cases, if you're using it for commercial purposes,

you might need to seek approval uh from the entity uh that created a third party uh package or capability that the skill is using. Um any any questions so far? Uh if there are any, please

um post them in the Q&A and we can answer them as kind of they arise. Okay. Um so moving forward uh one thing that we really kind of emphasize when creating cloud scientific skills um is

this interdisiplinary nature that I talked about um the fact that these skills encompass everything from bioinformatics to clinical research to machine learning to physics and quantum chemistry etc. And we we had um several suggestions um in terms of you know breaking them apart into subd

disciplines. So you know have a directory with only uh chemistry skills, another directory with

disciplines. So you know have a directory with only uh chemistry skills, another directory with only machine learning skills. But we decided actually not to do that because as we tested um these skills on kind of real real world scientific applications it became um pretty evident that the interdisiplinary nature of these skills is actually very important. Uh especially if you're

conducting novel research uh it's a very good idea that your AI assistant understands physics as well as data visualization as well as biological uh literature search. Um so that's why we kind of structure them as a as a kind of a standalone kind of single directory of skills to really encourage

people to um install all of them. Um and then um the I would say most of the publicly available databases, scientific databases are already covered in skills. Um everything from PubMed

um to uh Pubcam, Reacttoome, Geo, really kind of uh anything that comes to mind is probably already supported in skills. Uh but if not, please, you know, reach out to us on GitHub or our Slack and let us know. Um and then we can actually add uh the relevant um skills. I think there might be a

question. I believe we'll answer that in a second. Okay. Um now let's go into kind of the the reason

question. I believe we'll answer that in a second. Okay. Um now let's go into kind of the the reason uh you guys are here in terms of like how do you actually uh install install these skills. So

um let me go to our GitHub repository just to give you an idea of what these actually look like.

Um so the repository itself uh it has evolved a bit. Um again you know we we released this four months ago and that's why we called it cloud scientific skills. Um some of the instructions here uh for installation actually are specific to cloud code. Uh, but I'm going to give you a more

general um installation tutorial right now just to show you how to actually use it in any platform, not just cloud code. Um, the installation instructions here for cloud code are very clearly described. Uh, this was back in the day when skills could only be installed in

clearly described. Uh, this was back in the day when skills could only be installed in what cloud code called a plug-in. Uh but now now that has actually changed and I'm going to change to kind of show you what where what that change is. So um going back to kind of my

slide here um there's the the easiest way to install skills is you basically just download them. Um if you look at the the structure of the repository, it's actually very simple. It's

them. Um if you look at the the structure of the repository, it's actually very simple. It's

uh there's only a single directory here that you need to be concerned about called scientific skills. And here it has the list uh of directories for all the skills that uh cloud

scientific skills. And here it has the list uh of directories for all the skills that uh cloud scientific skills currently supports. So let's look at a quick example. Um let's look at maybe um what might be interesting here to the general audience. Um let's say polars. So

polar for those of you who are not familiar with it, it's uh Python packages uh similar to pandas uh but it's very computationally efficient. It allows you to work with uh data sets are very very large. So if you look at the polar skills uh you'll notice that there's uh you know one file

very large. So if you look at the polar skills uh you'll notice that there's uh you know one file as we had described earlier called skill.md and then some references. And if you go to skill.mmd,

you'll see that the skill has um a name uh a description um and a license associated with it. So the name of the skill is just polars in this case and the description is what the LM agent

it. So the name of the skill is just polars in this case and the description is what the LM agent uh reads to decide whether this skill should be used or not. So in this case, you know, this is a fast in-memory data data frame library for data sets, etc., etc. um you know it's used for lazy evaluation, parallel execution uh and so forth. Um and here you know we're

just giving the agent a bit more instruction that uh this is really useful for larger than memory uh data sets especially if you're running on a kind of a local laptop or something a bit more compute constrainted. And the skill itself again it has human readable instructions but these are

compute constrainted. And the skill itself again it has human readable instructions but these are really geared towards uh the agent. So this is your you're essentially telling the agent uh you know what this is, how to install it, uh how to run it, uh basic ideas behind it,

um and so on. So um the way you actually install a skill is uh you take the specific directory, let's say, you know, Astropi or CIRC or whatever, and you put it in the subsequent skills directory in whatever you're actually interested in using skills for. And let me explain what that means.

So I'm going to stop sharing uh for a second. I'm going to switch to uh another view. Um so

okay. So just sharing my entire screen here so everyone can actually see um my directory structure. Now I already have um I already have installed

um cursor and cloud code. So for those of you that haven't installed it already um we'll send out instructions for how to install cloud code and cursor. Uh

but honestly the instructions provided by cloud code and cursor are pretty clear. So you should be able to follow them directly. Uh but reach out if you have any questions. But what happens when you install cloud code and cursor in your home directory and this is a I'm I'm using a Mac here. Uh obviously if you're using uh uh Windows uh it's it's very similar. You're going to have

here. Uh obviously if you're using uh uh Windows uh it's it's very similar. You're going to have to kind of find where that directory is. But um if you look at u here there's something called uh where is it? Uh dot do.claude. This is a hidden directory because it starts with a

dot and inside.cloud is a folder that says skills. And essentially the way you install a skill is you uh grab it from whatever source including cloud scientific skills and you basically just put it in this directory. So again uh you can clone the repository or just download it. So I'm going to

just show you how to download it. That's the easiest thing to do. Um, I'm going to click on code. Um, and then, uh, download zip. And this is just going to zip it. Um, and download it.

code. Um, and then, uh, download zip. And this is just going to zip it. Um, and download it.

And then if you look at the contents, you'll find that it's basically just a bunch of folders. It's

very simple, a bunch of folders. And then if I want to um install a specific skill um I can just copy it. But what I'm going to do right now is actually I'm going to copy

all the skills and I'm going to paste them into the skills directory include cloud.

Okay. So now uh if I launch cloud code which I'm going to do right in front of you. Um I'm going to launch it from within cursor but please don't mistake

of you. Um I'm going to launch it from within cursor but please don't mistake this for cursor. I'm just using the terminal here. So if I launch claude uh and if I type uh slashkill this is going to now show me the list of all available skills.

So now cloud code has access to all the skills um I just installed. So these are all 141 scientific skills are now ready for cloud to use. Um and same thing actually goes for for cursor as well. So cursor um introduced support for skills I would say about two weeks ago and if you don't have the most up-to-date version of cursor you

probably need to update it. But if you go to the cursor settings and then if you go to rules, skills and sub aents uh here you can see all the list of installed skills and right now I have all the cloud scientific skills installed and the way you install them in cursor is very

similar to cloud code. Um again you go to uh you go to um dotcursor instead of cloud um dotcursor.

I hope everyone can see that. I know it's very faint. And there's a um directory um called um where do they store them now?

There should be a directory here um essentially called a skills. Um so by the way every um every creator of a uh certain client uh gets to decide exactly how skills work.

Um but essentially I would paste the skills here and these would become available to uh cursor. Uh now one thing to keep in mind is that whether you're in cursor or cloud code

uh cursor. Uh now one thing to keep in mind is that whether you're in cursor or cloud code uh you can actually have um skills that are specific to a certain project.

So for example, if I were to let's say I was working in my uh preferred IDE here, um for me it's it's cursor. Um I can actually create a new directory called uh cursor/skills and I can just place the skills I'm interested in for this project in the skills directory.

And these skills now uh can be used locally by the project. Um, and same goes uh for for cloud code as well. [clears throat] Okay. And before I open up for questions, um, if you want to learn how to install cloud code, um, you can just Google cloud code installation.

It's it's pretty straightforward. Um, if you want to use cursor, again, go to cursor.com. Uh,

you should be able to kind of easily download it and and install it. Uh but before I I go into Q&A, I do want to emphasize that, you know, in addition to our open source uh repository Cloud Scientific Skills, uh we do have a uh a fully hosted platform uh called K-Dense Web. Uh and what this is is

essentially it takes cloud scientific skills um integrates them and then builds on top of them um additional um infrastructure including compute infrastructure and additional capabilities. Uh and

it's fully hosted. So there's zero kind of setup uh required by the user and right now it's it's really kind of free free to use if you want to get started with it. So do do check it out. Um and this is kind of what what it looks like but this is not a you know webinar around

out. Um and this is kind of what what it looks like but this is not a you know webinar around K-Dense web. It's really focused around cloud scientific skills. And what I'm going to do

K-Dense web. It's really focused around cloud scientific skills. And what I'm going to do right now is I'm going to take a few questions um and then we'll go we'll hop back to cursor and I'll show you a quick demo for how to use uh cloud scientific skills and uh cloud code.

Um, let me unshare for a second just so I can see all the questions.

Okay. Um, so we have about eight questions.

Um so there's a question about if you add a lot of skills let's say you added 140 of these um cloud scientific skills does it eat into the token count? So yes but it's negligible. So

um let me share my screen here. So what actually gets loaded into the context um is is only uh let's look at an example. Let's look at um and data. This is used for single cell

sequencing analysis. If you look at the skill.md file, uh this is the only context that actually

sequencing analysis. If you look at the skill.md file, uh this is the only context that actually gets loaded into the LLM context. Um until the skill is needed. When the skill is needed, um this entire thing is loaded. Uh but if the skill is never used,

uh this is the only context that's actually loaded.

Um I hope I hope that kind of answered that. Uh, another question is um, does adding more skills uh, bias the LLM towards a specific skill? So, okay, this actually uh, is a is a broader

question. So, um, in the LLM context, in the agent context, there's all these descriptions

question. So, um, in the LLM context, in the agent context, there's all these descriptions of these skills. Now, that can cause a few issues sometimes. Let's say you had multiple skills that

had some contradictory information. For example, even in cloud scientific skills, there might be a um a PubMed skill that allows you to search PubMed. Uh but there might be another package, let's say GG get that also has capabilities internal to it that let you search PubMed. Um and

the description itself, there's two descriptions now for for two skills and they both allow you to search PubMed. So in these cases uh we were actually talking to the creators of skills from Entropic and they said what what tends to happen is um the model gets a bit confused. Um it doesn't

know whether it should call skill A or skill B. So when we design um the descriptions for these skills we're actually very careful in terms of making sure there's very little contradictory information between the skill descriptions. Um so if we think that the dedicated PubMed skill search

skill is better than whatever that Python package package offers in terms of search we make sure that PubMed is not mentioned in the description for gigg we only make sure it's mentioned for the PubMed database skill. And then a related question to this is um how do you make sure

that the models actually um invoke uh skills and this is still an unsolved problem unfortunately um you know we've we've talked to several model providers including Entropic um and um forcing

the agent to invoke a particular skill does need some let's say art. Uh when I say art, um there's no concrete uh best practices for how to make sure an agent always invokes the right skill and invokes it at all. Sometimes it doesn't invoke it. Like sometimes you might be using clot

code. It has the PubMed database skill and you're saying I want to do literature review on new

code. It has the PubMed database skill and you're saying I want to do literature review on new papers in single cell sequencing for disease X. Um and cloud code just fails to actually uh use the skill. So what we've noticed internally uh at K-Dense is the best way to invoke the skills is to

the skill. So what we've noticed internally uh at K-Dense is the best way to invoke the skills is to always be explicit to the agent and just tell it. Um there's two ways of telling it. You can either say, you know, the the kind of phrase we always like to use is use available skills and only

uh revert to kind of your own internal knowledge or your coding abilities when there's no skill that fulfills that requirement. And when we use the phrasing like this is actually helps the model understand that our first priority is for the agent to actually call the skills and then it can fall back on other mechanisms if it needs to. Um the other way to do it is to explicitly

mention the skill name. So for example in cloud scientific skills uh you know we might have a skill called um let's say there's a a skill for protocols IO integration. So protocols

io is u a website that essentially allows you to create these kind of interactive biological protocols. Uh if you want to use that skill you can actually just explicitly mention it

protocols. Uh if you want to use that skill you can actually just explicitly mention it in your prompt. You can say you know using the protocols IO integration skill do xyz and that in most cases you would probably get about 90% success in terms of invoking that specific skill.

Um, okay. Um, next question is, uh, would this workflow be the same for Windows users? Uh, yeah,

pretty much. I think the only difference between Windows and and Mac is, uh, where you find the cursor and the.claw, um, directories. But again, if you look at the documentation for clawed and cursor, etc., uh they have documentation for um uh Windows users, Linux users, uh Mac users, etc.

And they clearly kind of kind of outline where all of this stuff is. And you know, if if anyone you know does get into technical problems, please join our Slack community. Uh everyone in the company is very engaged and will happily answer any questions that you might have. and you know we'll be happy to also schedule kind of a a Google meet call um to give you a one-on-one tutorial as well. Okay,

another question is um is there a difference between cloud code and open code open code in terms of efficiency compatibility? Um I haven't personally used open uh open code. Um we primarily use uh cursor for our day-to-day tasks as well as um obviously um um cloud code. Uh open code

from my understanding is similar to cloud code uh but it lets you easily use nonattropic models. I

actually don't know how well skills are supported in open code but if anyone does uh you know please reach out and and let us know. Um another question is um are you testing models other than claude opus? I was wondering which open weight model is good at scientific skills. Uh we have tested a

opus? I was wondering which open weight model is good at scientific skills. Uh we have tested a few models. Um now keep in mind that the skills concept uh came out um before a lot of these

few models. Um now keep in mind that the skills concept uh came out um before a lot of these frontier models were released. What that basically means is during training and fine-tuning of these models they haven't been explicitly taught how to use skills. Now there are a few exceptions.

Uh Opus uh 4.6 and I believe also 4.5 is actually pretty good at using skills. Uh because as far as we're aware, Antropic during the kind of uh fine-tuning stage um taught it how to use skills um as a tool. So Opus works very very well with uh cloud scientific skills and other agent skills.

Uh Gemini 3 Pro also works pretty well. Uh you do have to nudge it a bit. you have to be a bit more explicit. Say use skills uh for you know X. Make sure you use available skills. Opus we find

more explicit. Say use skills uh for you know X. Make sure you use available skills. Opus we find you don't you don't have to do that too much. Uh solid 4.5 is also okay but not as good as uh as Opus. Now with regards to openweight models um we we have tried a few openweight models with

Opus. Now with regards to openweight models um we we have tried a few openweight models with cloud scientific skills uh including like Quen um I believe the openweight GPT model as well.

Um, frankly, they're not good. Um, a lot of these openw weight models are just not good uh at calling skills. I think they're generally not very good at instruction following. Um, and skills highly depend on how good the model is at, uh, instruction following. Uh, another question is um,

context management and memory are two big issues with cloud code. Is there anything in science skills related? If not, what do you suggest? So um in terms of context management, cloud scientific

skills related? If not, what do you suggest? So um in terms of context management, cloud scientific skills really rely on the clients to do the context management including cloud code and cursor etc. Uh now having said that I think cloud code in terms of context management honestly is the best

out there uh in terms of token efficiency. Uh I think open AI codeex might come a close second.

Uh but cloud cloud codes is pretty good pretty good at this. Uh now keep in mind that uh agent skills are actually very context efficient because um the model is only loading what it needs to load um versus for example when you use explicit tools or MCPS where everything is loaded into context and you quickly blow up context. Um so a general recommendation if context management is an issue

and memory is an issue um abandon all you know MCPs where you can abandon all tools where you can and just utilize skills uh as kind of the core intelligence of whatever you're you're building or using. Uh Peter I hope that answers your question if not please do follow up. Um another question

using. Uh Peter I hope that answers your question if not please do follow up. Um another question we have is um so um question is I like terminal a lot but my students uh do not like it that much.

Can people use scientific skills with cloud co-work or another non- terminal UI? Yes. Uh

so as of recently I would say uh most of these um desktop clients including cloud co-work now support skills. Um, again, specific installation instructions, honestly, they vary from client

support skills. Um, again, specific installation instructions, honestly, they vary from client to client, but every client now actually has um somewhere online um a getting started tutorial on how to actually install skills. Um, and obviously, you know, uh again to just kind of pitch here our

our own uh platform, u K-Dense Web U is a fully hosted kind of UI, very easy to use and it runs cloud scientific skills. So, they can always use that and it's it's free to use right now.

Uh another question is um is there a more uh sorry is this is a more general question but as cloud scientific skills can assist with scientific writing or even peer reviewing how do you address concerns about intellectual integrity? Okay so cloud scientific skills has a bunch of skills are focused on uh academic writing. Um so you can actually create a full uh ready to submit

manuscript uh include scientific skills including scientific schematics uh methodology discussion references um etc. Um how do you deal with intellectual integrity? Well um keep in mind you know all of these AI tools are really just tools right and you know integrity comes from the user

and and not from the AI. Um so uh with whether it's cloud code or or K-Dense web uh they provide you some with some pretty incredible capabilities. I mean I I have a you know I did my PhD in bioengineering uh a few years ago and you know writing a paper uh might have taken me what maybe

two three months and this is after I have the you know the data and all the kind of experimental results uh just the writing process the references the figures the uh validation the checks etc uh now with platforms like you know cloud scientific skills and K-Dense web and cloud code etc you can do it in a few days. Um but again integrity is really up to the the scientist itself. Um these

are AI tools. They they augment the scientist and they kind of support the scientist. Um okay. Um

another question is uh this is a question for uh K-Dense Web. Is there any plans for an enterprise offering of K-Dense uh K-Dense Web? Uh yes uh we do actually have an enterprise offering. Um this

is mostly geared towards kind of larger companies where we have custom integrations and we actually build custom skills for them as well uh if they want to. Um okay uh another question is does cloud scientific skills essentially uh perform all the functions similar to K-Dense web. Um I would say

generally no. Um, so cloud scientific skills has all the kind of raw capabilities that you might

generally no. Um, so cloud scientific skills has all the kind of raw capabilities that you might see in K-Dense web, but it does require a lot more handholding. Um, so with cloud scientific skills, you could technically achieve everything you can achieve in K-Dense web, but you've got to be sitting there with cloud code or cursor or whatever client you're using iterating back and

forth quite a bit. Uh, because it doesn't do end to end very well. Uh, you essentially have very to be very explicit. For example, you might say, "Okay, here's my data set. Do some uh exploration of the data set. Um then use this skill uh to create some some visualizations. Um then

uh use this skill to create a machine learning model based on this data." Um so you have to handhold it a lot. And K-Dense web um essentially you can say here's the data, here's my objective, I want a PDF, a 20 page PDF about your results, go and it will just execute everything.

uh but if you have you know if you have enough patience and the required compute etc you can probably do most of the things you do in K-Dense web include scientific skills but it becomes a lot harder though to do um okay and then there's a question of um how is um how is it different than

HCP um this is a question by Indrail sorry I don't actually know what HCP here is. I'm going to skip it for now. Um, another question. Uh, so actually, um, Alper brought up um, brings up a good point.

So, there was a recent comparison from a company called Versel, uh, between skills and what's called an agents.md. Agents.m MD is basically a fixed set of instructions that you give the, uh, your your agent uh, that's always in in context. H it works pretty well but um the keep in mind the

agent MD file um the entire thing is actually loaded into context. So if you have uh a very long uh des let's say if you were to load all the cloud scientific skills into an agent on MD file uh you'll immediately hit your context limit like that's not even uh possible. So agents.mmd are

good if you have like a very specific use case. Um you might have like a let's say two or three scientific workflows you might be interested in. You can put them in agents.mmd ignore skills. But

in most cases actually for the types of you know research that everyone here is doing agents.mmd

by itself just doesn't doesn't work. uh what is helpful in in in agents.mmd is you can say in agism.mmd you can say uh you know use your use available skills uh whenever I ask you to do some

in agism.mmd you can say uh you know use your use available skills uh whenever I ask you to do some sort of you know task or always use available skills so things like you know these explicit general instructions uh putting them in agents.mmd is actually a good idea uh another question wow we

have a lot of questions okay let's keep going um another question is um should we modify any of these skills to improve improve or specialize certain tasks. Uh do you want to see the updates or how would you like to be notified? Slack would be good. Uh I do like the range of topics covered.

So yes, so you know um cloud scientific skills is open source. It's available for the community and really the goal behind it is for the community to improve these skills. Um so you know we are uh putting out a lot of these skills but we're not experts in uh in all of these fields, right? Uh,

I mean, I wrote a bunch of um, uh, kind of quantum computation skills uh, the other day, but I'm I'm I'm not a I'm not a I I know nothing about quantum computers, right? Um, so we really rely on the community to actually improve these skills. So, if you have specific expertise uh,

around a certain topic, please do improve those skills. Submit a pull request. uh we'll review it, make sure that it's you know it's not compromising security or privacy or anything like that and then we'll we'll integrate it into the the main codebase so everyone can actually uh utilize it.

So yes, please please do contribute especially if you have you know expertise in a certain field that u is not very common that's that's very helpful. Um okay so going back to the HCP question so um Indranell is asking how is skills different from MCPS. Uh it's actually very different. So in

MCPs um there's a bunch of tools usually behind an MCP server. So MCP stands for model context protocol and it's this other other kind of concept that entropic also introduced last year. Uh but it has a bunch of tools. Each tool u has a description and there's a description of

year. Uh but it has a bunch of tools. Each tool u has a description and there's a description of all the input uh arguments as well as uh the uh description of what's returned from the tool. Now

u if you have a large number of tools uh all of these uh descriptions are actually loaded into the NLM context and they quickly blow up your context. So to give you an example uh with cloud scientific skills um technically each skill might expose probably hundreds of tools right because these are uh very kind of high level skills they use you know python packages with have which have a lot

of functions etc. So effectively, you know, cloud scientific skills is equivalent to possibly, you know, hundreds of thousands of actual tools. Um, and if you were to package these as an MCP server, uh, it's just not even possible. I mean, even, uh, putting, you know, 30 40 tools,

uh, as an MCP server and loading it into context in an LLM, uh, that hits your, uh, token limit immediately. um you know, let alone 140 kind of scientific skills. So skills are definitely

immediately. um you know, let alone 140 kind of scientific skills. So skills are definitely um way more advantageous um than MCPS when it comes to uh context management. Okay. Um so I think I covered most of the questions. Um actually there's there's one more. Uh can

we add new skills to your git? Absolutely. Um can you add new skills to the GitHub repo? Yes.

Um so we actually encourage users to follow uh let me give some clarifying instructions here. So

um adding skills to our cloud scientific skills is very easy but we do ask people to adhere um to the standard. So um if you go to agentskills.io IO this is the open standard um that the AI community

the standard. So um if you go to agentskills.io IO this is the open standard um that the AI community has agreed on in terms of what skills should look like. So specifically under specification um it tells you how to structure the skill. So as I had mentioned you know there's a high level

skill name and a skill skill.md file. Uh there's some metadata you need to add. So every skill has to have a name as well as a description um and some optional fields etc. But if you you know if whoever wants to contribute to cloud scientific skills if you follow the specifications in

agent skills um and then submit them to uh cloud scientific skills as a pull request uh we review them you know we review new pull requests almost every day uh and then we'll merge them into kind of the main skills uh repository and contributions are you know highly highly welcome. Um, okay.

Another question is where can I see a change log or updates to cloud scientific skills and K-Dense web and how can I make suggestions for new features? So, let's distinguish the the two. Um,

so change log and updates to cloud scientific skills, they're mostly in the um release notes. So

if you look at um if you look at um the releases here under GitHub um you'll see um changes uh mostly highle changes like sometimes there are certain improvements to like the wording of a description or we add another reference file etc. We might not necessarily include it here

as just too much detail. uh but on a high level uh you can get an idea of like everything that's being added and and changed in cloud scientific skills. Uh to answer the question about uh K-Dense web so again K-Dense web is our hosted platform uh K-Dense web gets updated almost every two days

uh with additional features and enhancements. Uh we do not have an official uh change log at this time. Uh but we do have a newsletter uh that goes out weekly uh where we indicate kind of the the major changes and improvement to K-Dense web. Um so if you are interested

in uh in receiving our K-Dense uh newsletter uh just sign up to K-Dense web and you'll be added to our newsletter. Um and of course if you want to unsubscribe you can unsubscribe at any time.

Now in terms of suggestions um so for cloud scientific skills if you have suggestions or improvements uh please go ahead and uh create an issue uh in GitHub um and that will kind of tell us to keep an eye on it and and uh kind of address it or or make that improvement. Also communicate

uh in Slack. So we have a a Slack community uh for K-Dense uh both for cloud scientific skills but also our other open source repositories. We have a bunch of them. Um, as well as K-Dense Web. Um,

if you have any feature requests, suggestions, improvements, please put them there and we'll we'll, you know, we'll we'll gladly um address them. Um, okay. I think that's all for questions.

Um, are there any questions around um how to best utilize cloud scientific skills? I know

we're almost at our 50-minute mark and this is supposed to be a 50-minute webinar so we can let people go. Um, but uh what we'll actually do is uh we want to record a few uh use cases. So let me

people go. Um, but uh what we'll actually do is uh we want to record a few uh use cases. So let me um share with you what I'm what I mean by that. So right now if you go to um cloud scientific skills

um you'll notice that there's a directory uh that says docs and under docs uh there's examples. Um

so here there are examples around um a bunch of different domains everything from genomics to infectious disease research to uh creating scientific illustrations and visualizations. And

we have explicit examples uh for how to use these skills in cloud code. So for example, this workflow here uh is uh uh discovery of novel eGFR inhibitors for lung cancer. And these are the skills that the workflow uses. And this is how to actually run the workflow. Again, you don't

have to copy and paste this entire thing in cloth code or cursor. You can just do it kind of one step at a time. Uh but this is just to give you an example for how to use it in in cloud code. So

uh what we're going to be doing over the next kind of few weeks is recording more videos for how to use cloud scientific skills as well as K-Dense web uh for a variety um of of use cases in a variety of clients. Uh now actually something I would love for people to comment on in in chat is

uh what are some of the clients that everyone is interested in? Um I think most people right now that are attending are probably using cloud code. Uh but if there are any other clients that we can run write instruction material for uh including examples and installations

um you know please let me know in the in the Q Q&A. uh if you want to uh elaborate a bit or uh in the chat as well. So uh windsurf it seems like that's one of them. Okay. Um I think windsurf

uh should technically have the same skills structure as um cursor. I believe uh GitHub copilot also has the same skills structure as cursor. So usually there's a skills directory uh and you just put your skills in there and they become accessible. Um so I think most of

these platforms right now most of these clients and just going back to u to uh my other tab here.

So uh most of these clients I believe uh let's see yeah I think cloud code cursor vs code github copilot gemini uh google adk recently etc they all have the same structure where uh essentially you uh put your skills in a skills directory and

then the client client can can discover it. Um, and then there's a quick question here about how do I subscribe to the newsletter. Uh, so we don't actually have right now a way to explicitly subscribe, but if you do join K-Dense Web, um, if you sign up to the platform, you'll get added to the newsletter. Uh, but we'll we'll add the capability on the website to subscribe

to the newsletter. Um, now the newsletter mostly focuses around K-Dense web. Uh, we don't have a lot of content on cloud scientific skills. uh but we'll we'll add some as well um content regarding cloud scientific skills. And one question by the way we we always get continuously as a company is

the difference between K-Dense web and uh claude scientific skills. So the uh to kind of answer that we wrote a quick uh blog post um which um you know if you are interested please please take

a look um that um clarifies the main difference. So um this one uh if you Google it you'll you can easily find it K-Dense web versus cloud scientific skills. Uh it really explains you know why as a company we're maintaining both why we're putting effort into both h and how each one can be used

for different different types of complexity uh when it comes to workflows with cloud scientific skills needing a lot more handholding uh versus K-Dense web being kind of this endtoend kind of autonomous researcher. Okay. Uh before I let everyone go um are there any more questions?

autonomous researcher. Okay. Uh before I let everyone go um are there any more questions?

Okay. Um I think we can call it a day. Um I I don't think we actually have time to run a live example because cloud code takes you know takes a bit of time. Um but uh if you are interested in specific use cases by the way please reach out to me and what we can do is we can take your use case

um and we can record a tutorial video around using clause scientific skills uh in kind of optimized fashion uh for that specific use case and we can kind of share it with the community. Um so do reach out to me um email uh is timothy.kassis@k-dense.ai

it's also on our website K-Dense.ai AI um and uh Slack community. Uh we're always online. Please

shoot us a message if you have any questions or suggestions for a tutorial as well. Um with that, um I'll let everyone go and the recording should be available to everyone that attended this meeting. Uh and we'll send it out to the attendee list, um as well. And we will be posting it on our YouTube channel um probably in a day or two uh if

you want to take a look back at some of the stuff that we talked about today. Thank you everyone.