Let’s return with the column Architectural Prompting: today it’s the section’s turn again The Technician’s Prompt. This time we change pace. Instead of a prompt to paste into chat, let’s talk about for the first time skillsa function to teach our working method to AI e call him back whenever needed.
As an example, I chose a frequent problem in writing professional content with AI: regulatory citations.
>> Do you want to stay updated and receive articles like this? Click here, it’s free
The problem: the quote that it seems correct
Technical reports, opinions, chapters of specifications, responses to a request for additions. More and more often they start from a draft written or revised with AI, and the drafts are full of references: an article of the Consolidated Construction Law, a paragraph of the Public Contracts Code, a technical standard, a sentence.
The trouble is, the assistant is very good at writing a quote that sounds perfectly believable. The number has the right format, the article seems relevant, the sentence has an organ and a date. Sometimes everything is correct, sometimes the number is wrong, the paragraph does not exist, the law has been modified or replaced, or the sentence does not say what it is made to say. And from the outside the two situations are disturbingly similar.
Whoever signs the document is responsible for what is written, even if the draft was written by a machine.
What is a skill
A skill is one folder which contains instructions for performing a specific task well, written in a text file called SKILL.md. Think of it as an operating manual or study procedure: it tells the assistant how to work, with which steps, with which rules and with which response format.
The file is made of two parts. At the top there is a presentation card, with the name of the skill and a description of when it should be used, then there are the actual instructions.
The operation has a peculiarity. The chatbot invokes the skill only when the user’s request matches the description. For this reason, many skills can be installed without burdening conversations: each comes into play only when needed.
Skills, prompts, system prompts: the differences
We write the prompt every time, for a defined task. If we change chats we have to start over.
The system prompt, which we talked about in the last article, configures a dedicated assistant, for example a customized GPT or a Project: excellent when we have a recurring task and we want an assistant who does just that.
The skill is portable. It is not tied to a dedicated assistant: we install it once and the assistant calls it on its own when it understands that our task requires it, in any conversation. We can also invoke her by name (“…for this task use the XYZ skill“). And it can be shared: it is a folder with a text file, so we can send it to a colleague like sending a document.
Where it can be used
Skills are now a feature present in most chatbots. They can be used in the app, on the web and on desktop, loading them from the settings section dedicated to skills or directly in the chat by asking the assistant to install it (“Hi Claude/ChatGPT/Gemini, install this skill”)
If, however, you use a specific assistant that does not yet have this function, no fear: the skill text also works very well as a system instruction for a custom agent or a project. It loses the convenience of automatic activation, but the method remains intact.
The skill I give you: verification of regulatory citations
This skill does only one thing, but it does it methodically. He takes a text, isolates every reference to laws, technical standards and sentences, and for each one he honestly says whether it has been verified at the source, whether it is inconsistent or whether it remains to be verified, indicating where to go to check.
The heart of the method is a principle that I have already talked about in the article on the anti-complacency prompt: the assistant must never agree with the text, and above all he must not confirm anything by heart. A quote is incorrect because it sounds plausible. If he cannot consult the source he writes it, and the reference remains “to be verified”. No judgments of convenience.
How to install:
- Open your favorite chatbot
- Transport the SKILL.md file into chat. which you can download below
- Waiting for the assistant’s ok.
- To use it, just write, for example: “Check the normative citations of the report I attach. The document is dated 30 September 2026″. The assistant will understand on his own that he needs to use the skill.
Instructions and suggestions for use
- Always indicate the date of the document to understand if the cited law was in force when the text was written and if it is still in force today.
- If you want a real verification and not a simple evaluation of plausibility, use an assistant with access to the web, or upload the texts of the rules yourself: the assistant can declare “verified” only what he has read from the source.
- UNI, EN and ISO technical standards often require payment. Without a copy they cannot be verified, and the skill will tell you that they remain to be verified.
A final note on skills
When a method works, it’s worth turning it into a skill. Your way of checking a calculation, reading an opinion, preparing a report can become a folder to share with the firm. It is the idea behind the principle that guides this column: The method before the tool.
Disclaimer: this skill is a support tool for the review of regulatory references and does not replace professional verification. The outcome of each check must be confirmed on official sources by the professional who signs the document, who remains solely responsible for the final content.
The weekly column “Architectural Prompting” is edited by experts Luciana Mastrolia, Giovanna Panucci and Andrea Tinazzo
>> If you are interested in these topics, also sign up to the free Linkedin Newsletter AI & Design for Technicians, we’ll talk about it here!
