Artifacts are something that Claude creates as a result of your chat as a separate file. For example, it might create a PowerPoint or a Word file, or even some Python code that it can execute.
A skill is a set of instructions that stay in Claude, and those instructions will execute whenever the right trigger phrases are used or you directly ask Claude to use the skill.
Artifacts are outputs — rendered files or interactive content (React apps, HTML pages, documents, code) that Claude produces and displays in a side panel during a chat. They’re the result of a conversation, meant for you to view, download, or use.
Skills are inputs — reusable instruction sets stored as files (SKILL.md format) that tell Claude how to behave for a particular task. Instead of explaining the same instructions every time, you store them in a Skill and pull them up whenever needed. Claude can even automatically recognize when a Skill is relevant to your task without you explicitly invoking it.
I built a personal AI operating system on top of Claude Skills and honestly I think it's changed how I work more than anything in the last five years, maybe more. And I say that as someone who's been skeptical of most of the hype. Your framing -- 'Claude recognizes the task and activates the right Skill' -- is exactly the shift that happened for me. I went from maintaining a prompt library I had to manually remember existed, to a system that just... surfaces the right behavior when I describe a problem. I don't invoke it. It finds me. The next unlock, for your audience, is chaining Skills so the output of one becomes the input to the next. That's where it stops feeling like a better search bar and starts feeling like something actually working alongside you.
I read a chapter from my English textbook when I was a kid, and there was this chapter called Mother's Day where a boy named Ruben gifted his mother a brooch, and when he was old, with 3 sons and 12 grandsons, the author wrote, Ruben's wife says, 'Ruben never changed from the sweet boy he once was."
I built up skills using the inbuilt skill creator and yes, the results are startling, I began to get more clarity and realistic responses that can be traced back to references, what I feel about a a need to build up a RAG kind of skill when I need to do a bit of research based on documents which I have on my local machine, feeding the document library to Claude will consume a lot of tokens so a RAG methodology would be very useful.
This is very helpful. I like your “teaching” style because you make sure everyone can easily follow what you’re to teach. I appreciate the effort you put into making your writing comprehensive and comprehensible. Thanks, Ruben.
Super helpful Ruben, I would love some pointers on how to set up an AI personal assistant agent to automate my day and week - things like managing my calendar, meeting preps and notes and emails within my work network. Any advice would be super helpful
great guide, really appreciate the honesty about the Max plan cost. that part gets skipped way too often. one thing I'm curious about though, has anyone actually tested the SKILL.md portability on Gemini or ChatGPT? would love to see that covered in a future post
makes sense, thanks for confirming. will keep an eye on whether other platforms adopt the format. really useful guide overall, looking forward to the next one
I realized that I want to a piece of my workflow in an existing skill to be broken out. Can I edit a skill to remove that piece, or do I need to create a new skill to replace the original? If I can edit it, are there best practices for that process?
you are incredible for all this. thank you so much! I'm currently working on building my voice document from your instructions... I hit my Claude limit after about 20 questions and need to wait another 6 hours each time. am I doing it wrong? anyone have advice? I'm a $20mo plan.
What stands out here is how quickly a process becomes invisible once it’s systematized.
At first, everything feels manual—thinking, prompting, refining. Then the structure gets embedded, and suddenly the output looks effortless, even though the work has just been moved beneath the surface.
Markets evolve the same way. What starts as discretionary decision-making eventually becomes rules, models, flows. And once that happens, price stops reflecting individual intent and starts reflecting the system itself.
By the time something feels “automatic,” it’s usually because the edge has already been packaged and distributed.
Yesterday ,I have encountered a scenario where in claude initiated a AI resume app,but failed to launch,I did RCA and I found that Authkey is the issue(CROCS).May be you can throw light on these kind of issues.
We spent three years learning how to talk to AI.
Prompt courses. Frameworks. "Act as a..." templates. We got really good at being translators between our brain and the AI's brain.
Skills just deleted that job.
You don't speak AI anymore. You just speak. Claude figures out which process you need based on how you talk.
a translator AND a broken record: same prompt, same context, every single chat.
now there’s no repeating yourself. Claude recognizes the task with one command :)
I found this super helpful but I am really confused as to how skills in cowork differs from anthropics own „Artifacts“?
Artifacts = the output - it's what Claude generates during a conversation.
a piece of code, document, or a chart. it lives inside that one chat. you close the chat, it's gone (unless you save it somewhere).
Skills = the instructions that shape the output - it's a reusable file (http://SKILL.md) that tells Claude how to do a task.
your process, steps, voice. and it fires automatically so you don't have to repeat prompts every time. it works across every conversation.
Artifacts are something that Claude creates as a result of your chat as a separate file. For example, it might create a PowerPoint or a Word file, or even some Python code that it can execute.
A skill is a set of instructions that stay in Claude, and those instructions will execute whenever the right trigger phrases are used or you directly ask Claude to use the skill.
I really wonder if you took your question and prompted Claude, what would be the answer?
Artifacts are outputs — rendered files or interactive content (React apps, HTML pages, documents, code) that Claude produces and displays in a side panel during a chat. They’re the result of a conversation, meant for you to view, download, or use.
Skills are inputs — reusable instruction sets stored as files (SKILL.md format) that tell Claude how to behave for a particular task. Instead of explaining the same instructions every time, you store them in a Skill and pull them up whenever needed. Claude can even automatically recognize when a Skill is relevant to your task without you explicitly invoking it.
I built a personal AI operating system on top of Claude Skills and honestly I think it's changed how I work more than anything in the last five years, maybe more. And I say that as someone who's been skeptical of most of the hype. Your framing -- 'Claude recognizes the task and activates the right Skill' -- is exactly the shift that happened for me. I went from maintaining a prompt library I had to manually remember existed, to a system that just... surfaces the right behavior when I describe a problem. I don't invoke it. It finds me. The next unlock, for your audience, is chaining Skills so the output of one becomes the input to the next. That's where it stops feeling like a better search bar and starts feeling like something actually working alongside you.
skeptics turned believers are always one of the most credible voices :)
i’ll see what i can do on chaining Skills - most haven’t built their own Skill yet.
Ruben never changed from the sweet boy he once was.
thanks? aha :)
I read a chapter from my English textbook when I was a kid, and there was this chapter called Mother's Day where a boy named Ruben gifted his mother a brooch, and when he was old, with 3 sons and 12 grandsons, the author wrote, Ruben's wife says, 'Ruben never changed from the sweet boy he once was."
What a wonderful story it is.
I built up skills using the inbuilt skill creator and yes, the results are startling, I began to get more clarity and realistic responses that can be traced back to references, what I feel about a a need to build up a RAG kind of skill when I need to do a bit of research based on documents which I have on my local machine, feeding the document library to Claude will consume a lot of tokens so a RAG methodology would be very useful.
that’s the power of structured context living inside the AI instead of you re-explaining everything every time :)
This is very helpful. I like your “teaching” style because you make sure everyone can easily follow what you’re to teach. I appreciate the effort you put into making your writing comprehensive and comprehensible. Thanks, Ruben.
that means a lot, seriously - thank you :)
i will write even better guides for everyone.
How can I save this guide please?
open the guide. hit the three dots (top right). click "Save for later."
if you want it offline, print to PDF from your browser :)
Super helpful Ruben, I would love some pointers on how to set up an AI personal assistant agent to automate my day and week - things like managing my calendar, meeting preps and notes and emails within my work network. Any advice would be super helpful
have you checked out Claude Computer Dispatch? i wrote a guide about it: https://ruben.substack.com/p/claude-computer
Ok this is great, thank you! Will check it out now and try it over the weekend
great guide, really appreciate the honesty about the Max plan cost. that part gets skipped way too often. one thing I'm curious about though, has anyone actually tested the SKILL.md portability on Gemini or ChatGPT? would love to see that covered in a future post
for now it’s best and works directly in Claude environment.
makes sense, thanks for confirming. will keep an eye on whether other platforms adopt the format. really useful guide overall, looking forward to the next one
I realized that I want to a piece of my workflow in an existing skill to be broken out. Can I edit a skill to remove that piece, or do I need to create a new skill to replace the original? If I can edit it, are there best practices for that process?
use Claude's Skill Creator - prompt it with something like:
Use the skill-creator to review and edit my existing [exact-skill-name] Skill.
Specifically:
Remove the [describe the exact piece] part of the instructions/workflow.
Keep all other sections, logic, tone, and structure exactly as they are.
Preserve the original skill name and directory/folder name unchanged (do not create a v2 or rename anything).
First, analyze the current http://SKILL.md (and any reference files), then output the updated skill folder ready to re-upload.
make sure to re-test the Skill.
Thank you.
Ruben—I have not received a welcome email (and the associated .md files). Can you provide this? Thanks
Hi - you just subscribed today so it takes a while to arrive.
if nothing: search for ruben@substack.com in your email to find my welcome email (also check spam/promotion).
make sure to click the “Step 1 - Get Prompt Library” button. it will send me an automatic email then you’ll receive my prompt library.
PS: the files are inside the prompt library - named “.md files”
Please give me the "anti-ai-writing" file
sent you a dm :)
you are incredible for all this. thank you so much! I'm currently working on building my voice document from your instructions... I hit my Claude limit after about 20 questions and need to wait another 6 hours each time. am I doing it wrong? anyone have advice? I'm a $20mo plan.
it’s why i upgraded to Max. i was hitting those limits too :)
What stands out here is how quickly a process becomes invisible once it’s systematized.
At first, everything feels manual—thinking, prompting, refining. Then the structure gets embedded, and suddenly the output looks effortless, even though the work has just been moved beneath the surface.
Markets evolve the same way. What starts as discretionary decision-making eventually becomes rules, models, flows. And once that happens, price stops reflecting individual intent and starts reflecting the system itself.
By the time something feels “automatic,” it’s usually because the edge has already been packaged and distributed.
when someone gets a great output it’s becase they spent time building the system.
Skills are the same. you front-load the thinking once. then it fires on its own. the heavy lifting is done.
@Ruben Hassid you again nailed it
thank you!! any ideas on what i should cover next? :)
Yesterday ,I have encountered a scenario where in claude initiated a AI resume app,but failed to launch,I did RCA and I found that Authkey is the issue(CROCS).May be you can throw light on these kind of issues.
Evolving very fast..
it really is - the important thing is we move just as fast.