Your AI Assistant Has Amnesia — One Simple File Can Fix It

Imagine a colleague with amnesia

Think about how you would teach a new colleague at work. On the first day, you explain: this is how we write invoices, this is the tone we use with clients, this is how we name files. If your colleague had amnesia and forgot everything overnight, you would explain it all again every single morning. That is exactly how a chatbot works today. Every conversation starts from zero.

Now imagine that instead, your colleague kept a small notebook. You write the instructions once, on a card, and your colleague reads the card every morning before starting work. That card is what the tech world calls a "skill" — a file that tells your AI assistant how to do something the way you like it, every single time.

The usual alternative is a "prompt" — the instruction you type into the chat. A prompt is like a sticky note: it works for that one conversation, and then it is gone. A skill is like a recipe card you keep in the kitchen drawer: you write it once, and you can use it forever.

What exactly is a skill?

A skill is a small folder with one important file inside, called SKILL.md. The name does not matter — what matters is what is in it: a short description of what the skill does, and then plain instructions, written in normal language. "When I ask you to write an email, start with a greeting, keep it under 150 words, and end with my signature." That is it. No programming required.

The idea comes from Agent Skills, an open format first released by Anthropic in December 2025. Because it is open, the same file works across many tools — from ChatGPT-style assistants to coding helpers used by developers. You write the skill once, and your assistant loads it whenever the task matches. The assistant does not read every skill you own at once; it only opens the one it needs, a bit like pulling one recipe card out of a box instead of emptying the whole box on the table.

Why this matters if you are not a programmer

Skills are not just for developers. They are for the small-business owner who sends the same type of invoice every week. For the teacher who prepares lesson plans in the same structure. For the volunteer who writes the same kind of newsletter for their club.

Here is how it works in practice. You open your AI assistant, and instead of typing the full instruction again, you tell it which skill to use. The assistant follows your saved instructions — the structure, the tone, the details you chose once. The first time, writing the skill takes ten minutes. After that, you save ten minutes every single time the task comes up. Small tasks become one-click routines.

The memory trick: four files instead of a database

Skills teach your assistant how to work. But what about remembering things about you — your clients, your projects, your preferences? A popular approach, shared by AI coach Alex Prompter, keeps it remarkably simple: four files, zero databases.

Instead of a complicated system, you keep one file per topic — one for people, one for companies, one for tasks. Then one small index file lists what each file contains, so the assistant knows where to look. It is like a filing cabinet with labeled folders instead of one giant pile of loose papers. Everything is readable, searchable, and easy to back up. If a file mirrors something outside — like your client list from a CRM — you write the date of the last update at the top, so the assistant knows how fresh the information is. Outdated information is how confident-sounding mistakes happen.

A European bonus: your instructions stay yours

For readers in the EU, there is an extra reason to like this approach: skills are files on your device or in your project, not in a company's database. Nobody trains a model on your private instructions. You can see every word, edit it, and delete it. That is exactly the kind of control the EU AI Act and GDPR push for — transparency and user control over how AI systems behave.

Because the format is open, you are not locked into one vendor. If you switch tools, your skills come with you — the way your recipe cards work in any kitchen. For small European businesses, that means investing time in skills is safe: the work you put in stays yours, regardless of which assistant becomes popular next year.

One thing to watch: not every skill is safe

Community skills are shared online, and most of them are fine. But a security study from February 2026 found that about a third of the public skills it examined had flaws, and some contained hidden instructions that could do harm. Treat a downloaded skill like a downloaded app: check who made it, read the file before using it, and test it on something unimportant first. If a skill asks for access to your passwords or sends data somewhere unexpected — do not use it.

How to start this week

The best first skill is small. Pick one task you repeat often — a weekly email, a document template, a list you always format the same way. Write down the steps in plain language, keep it under 30 lines, and save it as a skill. Use it for a week and see how it feels. From our own newsroom, we can confirm the habit changes how you work: once the assistant reliably follows a written routine, you stop repeating yourself and start trusting the process. Learning to build these routines is one of the most practical AI skills a non-technical person can pick up this year.

Your assistant is not forgetful because it is stupid. It is forgetful because nobody gave it a notebook. Give it one — and watch how much less explaining you have to do.