A visiting friend taught me how to create a custom skill in AI. It was an eye-opener.
Prompt engineering was a thing when AI chatbots first started. The quality of AI responses depended on what you included in your prompt. Companies reportedly paid as much as $300k in annual salary for prompt engineer positions.
AI has largely taken over that role. You can ask AI to write those prompts for itself. Or you can ask one AI to write prompts for another AI. The prompts that AI writes are often more detailed and effective than the prompts you write.
A custom skill goes along these lines. You can ask AI to do something and go back and forth a few rounds to fine-tune it, or you can ask AI to create instructions for itself on how to do that task. The result is a reusable custom skill. You trigger the saved skill when you need to perform the same task again. AI reads the instructions and does it the way you wanted.
Skill Examples
I’ve created several custom skills so far. Here are two examples:
Reformat Text
I have a tax calculator for retirees. It produces a report with all the inputs and results. When someone questions an unexpected result, I ask them to send me the report so I can see the full picture. However, Gmail strips all the line breaks and formatting in the report. It comes over as a big block of text, making it hard to decipher what’s going on.
I asked AI to create a skill to reformat the big block of text back to how it looks in the original report. I gave it samples of the report and the text block. AI did a first round. I reviewed the output and asked it to make some adjustments. AI wrote and saved detailed instructions for itself on how to do it. That became a custom skill.
When I receive another big block of text next time, I paste it into chat and say, “Please reformat this.” AI reads the saved instructions and does it. I don’t have to explain the background or my requirements again. AI already has that in the custom skill.
Prescreen Advisors
I maintain a directory of financial advisors who give advice but don’t manage investments. Some advisors start as advice-only but add assets under management after I include them in the directory. Some advisors stop accepting new clients when they reach full capacity. I need to review advisors in the directory periodically.
I asked AI to create a skill for this review. I told AI where to look and what to look for. I asked AI to give me a confidence level for whether an advisor still meets my requirements. AI wrote instructions for itself as another custom skill.
When I need to review an advisor, I identify the firm and ask, “Is this advisor advice-only?” AI loads the saved skill, does the research, and gives me a confidence level and why. I still review manually afterward, but AI’s prescreening has made the review process much easier.
AI may have created software code behind the scenes to perform these tasks. I don’t see the code or run the code. I only make my requests in the chat, as I do in other interactions with AI.
Create a Custom Skill
Creating a skill in AI is like teaching a junior employee. Do it once and hand off the job. It takes some back-and-forth to create it, but it will be easy when you trigger the skill next time.
I’m using Claude in its desktop app. Claude says both free and paid accounts have access to skills. I’m not sure what account levels in other AI services let you create skills.
When you create a skill in AI, start your chat with something like “Please create a skill,” and then describe what you need. Or you can say “Please use your skill-creator skill …” The skill-creator skill is a built-in meta-skill in AI. It knows how to create other skills.
Downloadable Skills
You can also download pre-made skills. I haven’t used any because it’s so easy to create custom skills to do exactly what I want.
If you explore downloadable skills, treat them as any other downloadable software. Only install ones from sources you trust. Skills are human-readable. Read what a skill does before enabling it.
Trigger a Skill
AI knows what skills it has. You usually don’t need a secret handshake to trigger a saved skill. It detects from your chat when it should apply a saved skill. Or you can mention the skill to give it a nudge.
Please use the reformatting skill on this tax calculator report.
You don’t have to use the exact name of the skill. Enough context will give AI clues for which skill to use.
Refine a Skill
Skills are updatable. Like teaching a junior employee, you don’t have to do everything in one go. If your requirements change, or if you see AI getting something wrong when it applies a skill, you can tell AI to update the skill with your new requirements or what mistakes to avoid. The saved skill will become better over time.
Delete a Skill
You can read the skill AI created or delete it when you don’t need it anymore. Different AI services save skills in different places. The Claude desktop app has them under Skills in Settings. Technically, you can also update a saved skill manually, but I would ask AI to update it through chat.
Keep It Read-Only
AI can make mistakes, just like a junior employee. I don’t give AI any credentials or let it make any changes. If it reformats my block of text poorly, I still have the original. If it flags an advisor incorrectly, I still review and confirm. I let it do the legwork and report back, but I keep the final say.
It’s a great value to have a “junior employee” available 24/7, taught to perform tasks to my specifications. I’m sold.
Learn the Nuts and Bolts
I put everything I use to manage my money in a book. My Financial Toolbox guides you to a clear course of action.

AIskeptic says
Great info. Think twice about using AI (Skills or otherwise) to automate actions in your financial accounts. For example, transferring funds, re-balancing portfolios, and other recurring tasks that require logging into an online financial account. If your AI makes a mistake, hallucinates, goes rogue, or other undesirable behavior, the financial institution will not reimburse you. The reason is because you explicitly gave permission to your AI agent to access your account – legally no different than if you had given permission to your neighbor to access your account.
Harry Sit says
That’s a great reminder. I’ll add it to the main post. I’m not giving AI any credentials or letting it take any actions. Reformatting text and doing research are both read-only.
Arun says
Harry,
Thanks for the useful post and the reply about not sharing credentials.
Is interaction with the A I agent end-to-end encrypted?
Is it safe if one uploads a brokerage statement for portfolio analysis or sorting investments into categories, etc.? Or, is it necessary to “sanitize” documents removing names and account numbers, etc.?
Thanks.
AIskeptic says
As I understand it, you should always “sanitize” documents by using redaction or other methods. There is absolutely no guarantee of privacy when using AI models. To greatly minimize (but not completely eliminate) privacy concerns, you can use a browser that prioritizes privacy, and provides an anonymous front-end to your chosen AI model that ultimately receives your query (e.g. Claude Sonnet). One example of that arrangement is the Brave browser, which incorporates the Leo AI front-end into the browser. Using Brave + Leo AI + VPN will greatly reduce the exposure of your personally identifiable information (PII). By doing this, you are essentially inserting a virtual layer between your PII and Big Tech, thereby changing the trust arrangement from you trusting Big Tech, to instead trusting the provider of the virtual layer. But even with that arrangement, you should not include personal information in the query/prompt that you issue into AI.
As the tech that surrounds us grows in complexity over time, the ease with which Big Tech can deceive and con people grows also. This is because along with the growing tech complexity comes diminishing understanding by the average person. If people don’t understand the tech, then exploiting their ignorance becomes increasingly easy. Rest assured – this principle is at the very heart of the Big Tech business model. There are two ways to mitigate this exploitation in your own life – both require ongoing, deliberate, conscious effort. (a) be very selective about adopting tech, (b)for the tech you choose to adopt, understand its perils. Ironically, tech is touted to save us time and effort, but it has an equal but opposite cost associated with (a) and (b) above.
Caveat technologiae adopter (let the tech adopter beware) – LOL.
Harry Sit says
I would use the need-to-know principle. It doesn’t matter what encryption it has. There’s no need to disclose personal information only to analyze investments. Even a human advisor doesn’t need to know the exact names on the accounts or the account numbers. It’s easy to supply screenshots from statements or the positions pages without personal information.
Space Doc says
Spiffy. You are learning Harry. Good job. 😉
I use Ai to evaluate purchases of shares in farmland. It looks up historical info for farmland sales, where the farm gets its water from, climate predictions for the area, then delves into the profitability of the deals itself, projected returns, etc. It knows that I favor farms that have a good water source in an area that will not be negatively impacted by climate change (these are usually 8 to 10 year holdings, so that’s relevant), and with favorable terms. It’s actually a project in Claude with different skills applied within. I just feed it all the documentation, then it sorts it out and gives me a pdf report with recommendations. It’s getting better and better over time. Love it.
Dave S. says
“I’m using Claude in its desktop app”
—
So are you using the free account or the paid one?
Can you please clarify?
Harry Sit says
I paid $200 a year for a Pro account because I also need Claude Code. Claude says free accounts also have access to the skills feature highlighted in this post.