What's no longer worth learning

Going back over what I've published, I noticed that several articles ended up in the same place: with something I learnt with a lot of effort that was no use to me a few months later. I've gathered that drawer into a single article, with what replaced each thing and a short list of what I do think deserves your time today.

A couple of years ago I spent several weeks learning to chain instructions with a very specific structure, with roles, delimiters and "magic" words that supposedly made the model think better. I ended up with a twenty-page document full of templates. Today I don't use a single one. Not because they were wrong, but because the models changed and stopped needing them.

That feeling of having studied for an exam that got cancelled has happened to me quite a few times. And since this blog was born so I wouldn't forget what I learn, it also helps me not forget what I've unlearnt.

A note of humility before going on: this is what has stopped being useful to me, working mostly with Claude and also with OpenAI, DeepSeek, GLM, Kimi and MiniMax. It's not a universal ruling. If something on this list still works for you, great.

Prompt engineering is the technique of writing the requests you give an AI so it answers with what you need. A few years ago it made a lot of sense as a separate discipline: models understood less, got lost with long requests, and there were tricks that genuinely changed the result.

That has changed. Today a clear request, written in plain language, works almost every time. What's left of value from that discipline fits in a few lines, and it's the same thing you'd do with a new colleague:

• Give it context: who you are, what it's for, what happened before.
• Tell it the goal: what you want to achieve, not just what you want it to do.
• Give it an example if you have one.
• Tell it what format you want the result in.

And if it doesn't come out right, correct it by talking, just like with a person.

A before-and-after example. I used to write something like this: "Act as a marketing expert with 20 years of experience. Think step by step. Task: write an email. Constraints: maximum 150 words. Tone: professional. Format: ###SUBJECT### ###BODY###". Today I ask for the same thing like this: "I need to write to a client who hasn't replied to me about a quote for two weeks. I don't want to sound pushy. Write me a short email, with a subject line, in the tone I'd use with someone I've already worked with". The second one comes out better, and on top of that I understand it when I reread it.

I go into it more calmly in Asking AI for things without magic formulas.

They're sold everywhere: "500 prompts for your business", "the secret prompts the experts use". They're not worth it, for two reasons. They're texts you could write yourself in a minute, and they're designed for a specific model that will behave differently in a few months.

I downloaded a free one a while ago, in exchange for my email. It was about a hundred generic templates like "write an article about TOPIC". I used two. The email address, on the other hand, is still getting offers.

What replaces them is much more boring and much more useful: your own saved instructions for your tasks, written with your context. I have a small file with the instructions I repeat often, like how I want meeting summaries or how I review a text before publishing it. They're worthless to anyone else, and that's the whole point.

For a while I built a lot of automations with n8n, a tool where you draw the steps of a task one by one: an email arrives, it's read, a piece of data is extracted, it's saved to a sheet. It works, but every workflow has to be built, tested and maintained.

Now I use n8n a lot less, because many of those things I ask of an agent directly: an AI that takes several steps on its own and is connected to my tools. Instead of drawing the workflow, I explain the task.

A concrete example: I had a workflow that collected sign-up emails from new clients, pulled out the name, email and service purchased, and logged them in a spreadsheet. It took me an afternoon to build and it broke every time the email changed format. Today I ask the agent to go through those emails and update the sheet, and it adapts to format changes on its own. I gained flexibility. I lost some predictability: every now and then, the agent interprets a field differently.

Workflows still make sense for fixed tasks that repeat daily and have to come out exactly the same every time. For example, the scheduled backups of my servers or the alert when a service goes down: there I don't want creativity, I want the same thing to happen every day.

I explain it in full in Why I no longer need n8n.

Here I want to be fair. There's serious training, made by people who know their stuff and keep what they teach up to date. And there are courses that package what the AI itself would explain to you for free if you asked, with a pretty cover and a countdown timer.

Deciding whether an AI course is worth it isn't easy, because many look alike in the title but not in the content. These are the signs that make me suspicious:

• They promise income.
• They sell instructions or "secret prompts".
• They don't say when the syllabus was last updated.
• They teach a single tool as if it were going to last forever.

And the question I ask myself before paying for any training is very simple: what will I be able to do when I finish that I can't do today? If the answer is vague, the course probably is too.

I paid for one a while ago, one of those promising to master AI for business in a weekend. Of the twelve hours of video, about nine were demonstrations of a tool whose interface changed the following month. The useful part would have fit in an afternoon, and I already knew a good chunk of it. I'm not going to say which one it was; this isn't about pointing fingers but about learning to look.

What no course replaces is wrestling with a real case and getting it wrong a couple of times.

The middleman tools in the subtitle are those interfaces, extensions or platforms that wrap a model and add buttons to it. Some are great. The problem is learning them as if they were going to last for years.

I invested quite a bit of time in a browser extension that connected the chat to my documents. I learnt its shortcuts, its menus and its tricks. A few months later, the assistant itself built that feature in as standard and the extension stopped being updated.

What stays when the tool changes is judgment, not menus. Knowing what you want to achieve, how to check it and where the risk lies serves you with today's tool and with next year's.

This is the useful part. If someone asks me where to start to learn AI, I'd give them this short list:

1. Explaining clearly what you want, as you would to a new colleague. It's the foundation of everything else, and I cover it in Asking AI for things without magic formulas.
2. Checking the results and spotting when the machine makes something up, what's called "hallucinating". A text that sounds good isn't a true text.
3. Knowing what data never goes into any chat: client data, passwords, anything that isn't yours. I deal with it in Is what I do with AI legal?.
4. Understanding what an agent is and what its problems are: permissions, errors and costs. It's in The problems of working with agents.
5. Connecting AI to your tools, which is where it really starts saving you time. I tell it in How I connected AI to my tools.
6. Knowing what surrounds the model so it's actually useful: interfaces, your own documents, connections. I explain it in What surrounds the model to make it useful.

And one more, the one that needs no link: using AI on real tasks in your work. That's where you really learn. No syllabus teaches you what you learn the first time the assistant hands you something that looks perfect and is wrong.

To finish, the questions I ask myself before putting time or money into anything about AI:

• When was it last updated?
• Does it depend on a specific tool that could change tomorrow?
• Does it promise financial results?
• Could I get the same thing by asking the assistant itself?
• Does the person selling it show real cases, with their failures?

If two or more fail, I'm wary.

The drawer of discarded things will keep filling up, I'm sure of that. And I wouldn't be surprised if, a year from now, this very article ends up inside it.