Join Table for One | Other Articles
I grew up watching Rosey from The Jetson’s cook, clean, fold the laundry, and have a whole personality while she was at it. (I don’t know how niche that reference is but she was a robot maid on a cartoon).
But what I actually got is a chat window that can write a 900-word essay on how to correctly fold a fitted sheet and then wait for me to go do it myself. So while I wait for AI to give me the actual help I need… I will have to make do with what it can do right now.
I use claude skills to run almost everything in my life. They analyze my Substack. Run my job search. Read my inbox every morning. Plan my meals. Watch specific industries for news.
You get the picture.
A few of them run continuously on their own, a few act when a file changes, and a few of them sit there waiting for me to call on them.
Building a skill is easy but building a good skill is nowhere close to it.
The first few that I built could not give me repeatable results. But the recent ones not only give perfect results but they learn with every run and even work with other skills.
P.S. I am calling skills working with other skills as Tasks and this is my petition to Anthropic or Open AI to use let me coin this term!
Building Better AI Helpers
So what is a skill? By definition, a skill is a folder with a plain text file in it, and that file tells Claude/ChatGPT how to do one specific job the same way every single time. You write the instructions in ordinary English and save it in a SKILL.md file. Your AI reads it and any other documents you reference to execute tasks on your behalf.
I am a religious Claude user but whatever I explain below works with ChatGPT too.
Do Before You Build
This part isn’t just mine. It comes directly from Anthropic’s guide to building skills for Claude.
Complete at least one run of the task in a simple chat.
Work through the problem with your AI using simple prompts and as you go, you will notice that you naturally hand over context, explain your preferences, and describe the workflow. You have to keep track of what information you repeatedly provide to the chat. That should become your context files.
What works best for me is to do the task 2 to 3 times and keep a note of every single thing I had to correct. Let me walk you through how I built the meal planner skill.
Before I even thought of a skill, I started a chat and told it this:
Okay I need your help to plan my meals and create a grocery list, every Saturday or Sunday of the week, my main goal is to eat healthier, stop ordering out so much and actually save some money along the way. So I love cooking, but I am looking for recipes that are relatively easy to make while still feeling kind of fancy. And I will drop some links to my favorite recipes at the end of this message so you can understand my taste and suggest something along the line.
So it’s just me and my husband at home. I work from home, so I eat three meals a day, uh, six days a week, and my husband eats three meals a day on Mondays, Fridays, and Saturdays. Um, he’s also there for breakfast and all days except Sunday, but he usually prefers a smoothie, like a banana smoothie with protein powder.
I eat relatively smaller portions, so two slices of bread and two eggs with coffee is a full breakfast for me. This kind of keeps me full for at least four to five hours. And my husband eats the normal portion size.
So what you have to do is you have to plan the whole week of meals, uh, with me and then build a grocery list. But you also have to keep track of what’s already there in the pantry and in my fridge because I don’t want to waste any food. Um, so right now, I have a loaf of bread, maybe half a carton of milk, about five eggs, all of the spices, a box of pasta, coffee, and a block of cheese. Other than that, I have some spinach left, but, nothing else.
And my husband and I, we don’t have any food restrictions or allergies. We both love meat and seafood, but I don’t want recipes that are heavy on red meat and don’t overdo the meat selection. So don’t ask me to buy a pack of chicken, a pack of minced beef, and fish for just one week. Cause I usually marinate whatever food that I get when I get the groceries home. So just having, like, one or two meat and seafood pairing works better because it helps me prep my meals faster, and then I can just prep the entire batch of chicken for the week. Or if it’s shrimp, I can just prep the entire batch for the week so that so that I can optimize my prepping and my cooking time. I can maybe give about thirty to thirty five minutes of cook time a day, and I’m definitely going to cook two meals at a time. So if I cook for lunch, it’s gonna be lunch and dinner that I cook, and if I cook for dinner, I will cook for that meal and my lunch for the next day.
So it’s just me and my husband at home. I work from home, so I eat three meals a day, uh, six days a week, and my husband eats three meals a day on Mondays, Fridays, and Saturdays. Um, he’s also there for breakfast and all days except Sunday, but he usually prefers a smoothie, like a banana smoothie with protein powder.
I eat relatively smaller portions, so two slices of bread and two eggs with coffee is a full breakfast for me. This kind of keeps me full for at least four to five hours. And my husband eats the normal portion size.
So what you have to do is you have to plan the whole week of meals, uh, with me and then build a grocery list. But you also have to keep track of what’s already there in the pantry and in my fridge because I don’t want to waste any food. Um, so right now, I have a loaf of bread, maybe half a carton of milk, about five eggs, all of the spices, a box of pasta, coffee, and a block of cheese. Other than that, I have some spinach left, but, nothing else.
And my husband and I, we don’t have any food restrictions or allergies. We both love meat and seafood, but I don’t want recipes that are heavy on red meat and don’t overdo the meat selection. So don’t ask me to buy a pack of chicken, a pack of minced beef, and fish for just one week. Cause I usually marinate whatever food that I get when I get the groceries home. So just having, like, one or two meat and seafood pairing works better because it helps me prep my meals faster, and then I can just prep the entire batch of chicken for the week. Or if it’s shrimp, I can just prep the entire batch for the week so that so that I can optimize my prepping and my cooking time. I can maybe give about thirty to thirty five minutes of cook time a day, and I’m definitely going to cook two meals at a time. So if I cook for lunch, it’s gonna be lunch and dinner that I cook, and if I cook for dinner, I will cook for that meal and my lunch for the next day.
Just remember to not suggest recipes that are too elaborate.
Okay. So here are the few recipes that I have followed, and I have liked so that you know what kind of food I generally gravitate towards.
1. https://www.recipetineats.com/tomato-burrata-salad/
2. https://www.onceuponachef.com/recipes/perfectly-grilled-chicken-breasts.html
3. https://www.yourhomebasedmom.com/smash-burger/
4. https://www.recipetineats.com/marinated-grilled-salmon/
5. https://www.skinnytaste.com/healthy-breaded-shrimp/
I love eating spicy food. Um, so if you have spicy recipes, I would love that a lot.
One last thing, you have to make sure that I get my portion of vitamins and nutrients through the food. So when you’re making the grocery list, make sure you add some fruits like blueberries, oranges, and also veggies like spinach, beetroot, any other green leafy veggies that you can think of.
Before we start planning, you need to figure out what documents or files will we need to make this work. So from what I see, we definitely need a weekly plan log so that we know what we planned on week one, week two, etcetera, etcetera. We might also need a tracker for the pantry items. Uh, we might need a way for me to give you feedback on the recipes, um, that I liked versus that I didn’t like while we are brainstorming the recipes and also after I have cooked with them. So maybe at the start of every new session, after the first session, we can start with asking for my feedback on what happened last week, and we could also use this time to sort of get an inventory check of whatever I have in my pantry at the moment.
So did you understand the assignment? Do you have any questions? Make sure you do not assume anything if you have a doubt, just ask me the doubts right now.
Just remember to not suggest recipes that are too elaborate.
Okay. So here are the few recipes that I have followed, and I have liked so that you know what kind of food I generally gravitate towards.
1. https://www.recipetineats.com/tomato-burrata-salad/
2. https://www.onceuponachef.com/recipes/perfectly-grilled-chicken-breasts.html
3. https://www.yourhomebasedmom.com/smash-burger/
4. https://www.recipetineats.com/marinated-grilled-salmon/
5. https://www.skinnytaste.com/healthy-breaded-shrimp/
I love eating spicy food. Um, so if you have spicy recipes, I would love that a lot.
One last thing, you have to make sure that I get my portion of vitamins and nutrients through the food. So when you’re making the grocery list, make sure you add some fruits like blueberries, oranges, and also veggies like spinach, beetroot, any other green leafy veggies that you can think of.
Before we start planning, you need to figure out what documents or files will we need to make this work. So from what I see, we definitely need a weekly plan log so that we know what we planned on week one, week two, etcetera, etcetera. We might also need a tracker for the pantry items. Uh, we might need a way for me to give you feedback on the recipes, um, that I liked versus that I didn’t like while we are brainstorming the recipes and also after I have cooked with them. So maybe at the start of every new session, after the first session, we can start with asking for my feedback on what happened last week, and we could also use this time to sort of get an inventory check of whatever I have in my pantry at the moment.
So did you understand the assignment? Do you have any questions? Make sure you do not assume anything if you have a doubt, just ask me the doubts right now.
Then I ran this process on the same chat for 2 runs. Two runs of feedback, corrections, and approvals later, I had a decent understanding of what I needed to make this into a skill.
Design The System
Designing the skill is more than telling Claude what you want it to do. Designing means understanding the context required for this skill to perform, building a workflow for proper execution, identifying conditional dependencies, and cutting steps across the various commands.
A skill need not be a single entity that only does one thing. It can be complex.
For instance, I have a skill that edits my substack articles after I write them. But it is split into 3 commands: /substack-article-revise, /substack-article-structure-edit, /substack-article-final-edit. All three have different functionalities and require different context.
The logic for your design should come straight out of your runs. What did you correct? What did you have to repeat? Where did the chat go wrong?
Those should act as your specifications. The things that you have to repeat become context files, the things you correct become rules, and the things where the AI gave you a wrong output confidently become examples.
When I was use the meal planner chat, Claude kept on suggesting improper quantities. It kept finding recipes that were elaborate in ingredients and time. It couldn’t update what was in my pantry as easily. It ignored cost.
Every one of those issues helped me create the right files. Quantity mismatch led me to create a detailed portion size file that tracked what I actually used, how much it made, and how long did it last. I also created a live pantry file that updated when I accepted a recipe. The last one was a feedback loop that allowed me to say “the miso salmon was good but it took 60 minutes instead of 40.” So the skill would get better with every use.
But every skill you design doesn't have to be elaborate and doesn't need the same amount of freedom.
Anthropic frames the process of designing a skill as managing the degrees of freedom. How much depth and freedom you need to give to your skill depends on the expected outcome.
If you can achieve the same outcome via multiple routes, then you need to give the skill just some direction and a lot of freedom. But if the outcome has a serious chance of failure, then you need depth. You need guardrails, workflows and even examples.
Built It Elsewhere
Most advice I see on the internet for creating a skill is to use the same chat window to create and “package” your skill. And that is some of the worst advice I have ever heard.
The last time I built a skill inside the same chat window, the one that I had been using as a test ground, it backfired! Hard. That skill inherited almost every wrong action and decision, making it impossible for it to deliver the results I wanted.
So now I split the actual building into two steps. First, I go to the original chat window for a brief. I ask it to share the outcome we want, the decisions we made, the information we used, and the feedback we gathered along the way. Second, I open a new window and share my design, workflow, context files, rules, and examples (basically everything we created previously) along with the brief.
This helps me build a clean file that gets all the correct conclusions but not the detours.
Best Practices
The structure I just shared gets you a skill that works. Now the best practices below are the ones that unlock GOD MODE!
Order The Actions
Some skills have a sequence that they need to follow and some don’t. The meal planner has no sequential workflow as such. But my Substack skill and my Reddit skill do, and they have to run specific tasks in a specific order.
For the Reddit skill, it has to go through setup, daily scan, and a maintenance pass. In that order. Because if the skill runs them out of order then the output is worthless.
So when you come across a similar situation, write that order in the file itself as named phases or steps with rules for each and every phase. And make sure you define how the skill can know that an action is finished.
Keep A Log
Once you run something enough number of times, you naturally start to notice flaws. You think of better ways, easier ways to execute the same thing.
Version control is the single most important thing that helps me upgrade a skill without breaking it.
It might seem like an overkill to have “version: 2.3.1” for something that is used just by you. But, trust me, when your skill starts behaving differently and you can’t remember what you changed, version control literally saves you.
Build A Loop
Every skill should have a feedback loop. Not something you notice and then fix but something the skill notices and fixes.
Basically you need the skill to run the check, fix what failed, and run again to test. This, in my opinion, is the entire difference between a skill built by an amateur and one built by a professional.
Most of my skills now can run regression analysis on themselves. They have test cases, a matrix of good vs bad output, and a file to keep track of what was tested, what was changed, and what was the result.
A skill alone if fine. But somewhere in the last month I stopped building single skills. Now, I build Tasks. A task, the way I am describing it, is a collection of two or more skills/agents working with each other. They hand work to one another and even trigger each other when a particular action is taken.
I’ll share how I go about building those soon.
Until then, tell me how many skills/agents you use? Do you have any favorite ones?
Building the actual skill in a fresh chat window instead of the one you used to test it is advice that goes against almost everything else I've read about this. Inheriting every wrong decision from the test ground makes obvious sense in hindsight, but I would have packaged it in-place without a second thought.
I built an internal agent last year that I was convinced was "insanely good" because it nailed every test case I threw at it. Then I handed it to someone else on the team and it fell apart in about ten minutes, turns out I'd unconsciously trained my prompts around my own workflow quirks. Curious if you've hit that same wall: does "insanely good" hold up when someone other than the builder is driving it?