All Things AI · Season 3 · What You Can Build · Episode 16
A prompt is a request you type. A skill is a saved procedure that runs the same way every time. A bot is something that runs without you starting it. These get used interchangeably and they are not the same thing, and knowing the difference tells you whether a tool is offering you anything.
Why the words matter
Every product in this space describes itself with all three terms, usually in the same sentence. An AI bot with smart skills and optimised prompts is a phrase that means nothing and appears everywhere.
The consequence is that agents cannot tell the difference between a tool that saves a text snippet and one that genuinely runs work on its own, because both are marketed identically.
Thirty minutes on the vocabulary is the best-value time in this season, because it turns every future evaluation into a short conversation. What actually runs, what starts it, and what happens when it fails are three questions that cut through nearly all of it.
A prompt
A request you type. That is the whole definition.
It exists for one moment and then it is gone. If it produced something good, the good thing came from how you phrased it, and if you did not save the phrasing, you will not reproduce it reliably.
What it is genuinely good for
Anything one-off, exploratory, or unrepeatable. The majority of daily use is this and there is nothing wrong with that.
Where it stops working
The moment you need the same result twice. Two people typing the same intent get different outputs, and so does the same person on a different day, because the phrasing is doing invisible work and the phrasing changed.
A skill
A saved procedure. The instructions, the steps, the rules, and the format, written down once and reused.
The important part is not that it is saved. It is that it produces consistent output, because the variability was in the phrasing and the phrasing is now fixed. This is the difference between a result you got once and a result you can rely on.
It encodes decisions, not just words
A good skill holds everything you worked out the hard way: what to include, what to never do, what the output must look like, what to do when something is missing. It is institutional knowledge in a file, and it keeps working when you have forgotten why you made those choices.
It still needs you to start it
This is the boundary. A skill is a procedure waiting to be run. You decide when. That is the difference between a skill and the next thing.
It is where most of the value actually is
Not the exciting category, and the one that changes the most. Ten reliable skills for the ten things you do repeatedly is a bigger practical gain than any amount of autonomy, and it carries far less risk.
A bot
Something that runs without you starting it. On a schedule, or triggered by an event, or in response to something arriving.
The distinction is entirely about who starts it, and it is a much bigger distinction than it sounds. Something that runs on its own is something you are not watching, which means it needs to handle failure gracefully, know its limits, and be safe when the unexpected arrives, because you will not be there when it does.
It needs a defined boundary
What it may do, what it must never do, and what it does when it is unsure. An unattended process with no stated limits will eventually meet a situation nobody anticipated and handle it confidently.
It needs to be able to stop
Failing loudly and doing nothing is correct behaviour. Continuing while producing wrong results is the failure that costs weeks, because nobody notices until the damage has accumulated.
It needs checking even when it reports success
A process that says it succeeded is reporting that it finished, not that it was right. Anything running unattended needs a periodic look at what it actually produced.
The order to build them in
Prompt until you notice you are typing the same thing repeatedly. That noticing is the signal, and it usually takes a couple of weeks.
Turn it into a skill. Write down what you were actually asking for, including the parts you had been supplying implicitly by knowing your own business. This step is where most of the thinking happens.
Run the skill by hand for a few weeks. This is the step people skip and it is the one that matters, because it is where you discover what it does with the awkward cases.
Only then automate it, and only if the frequency justifies it. A skill you run twice a month is fine as a skill forever.
How to evaluate a product with this
Three questions, and most marketing does not survive them.
What exactly does it run, and can I see the instructions. If the procedure is hidden, you are buying a black box and you cannot fix it when it is wrong.
What starts it, and can I stop it. If it runs on its own, what are its limits and what does it do when it is uncertain.
What happens when it fails, and how would I find out. A tool with no answer to this is a tool that will fail silently, and silent failure in a business system is worse than no system.
Where this goes wrong
Three failures, and the first is the expensive one.
Automating before the procedure is proven. A skill that has not been run by hand for a few weeks has not met its awkward cases yet, and automating it means meeting them unattended. Almost every automation disaster traces back to this.
Building a bot for something that happens rarely. If it runs twice a month, starting it yourself takes four seconds a month and the automation is a system you now have to maintain. Automate frequency, not importance.
Keeping the procedure in your head. A skill that only exists as something you know how to ask for is not a skill, it is a habit, and it disappears the moment you are busy or unavailable.
What it grows
Every other episode in this season, which is why this one is here.
Once the vocabulary is clear, the rest of it stops being a collection of impressive-sounding possibilities and becomes a set of decisions you can actually make. Is this a prompt, a skill, or a bot. Have I proven it by hand. Do I know what it does when it fails.
Those three questions are most of the judgment required to run any of this well, and they are considerably more valuable than knowing which model is currently best.
What it costs
Nothing. This is the one episode with no tool, no subscription, and no setup. It is half an hour of vocabulary that makes every subsequent decision faster and most sales pitches transparent.
Questions agents actually type
What is the difference between an AI prompt, a skill, and a bot?
A prompt is a request you type once. A skill is a saved procedure that produces consistent output because the phrasing is fixed. A bot runs without you starting it, on a schedule or a trigger. The difference between a skill and a bot is entirely about who starts it.
When should I turn a prompt into a skill?
When you notice you are typing the same thing repeatedly, which usually takes a couple of weeks. Write down what you were actually asking for, including the parts you supplied implicitly by knowing your own business, because that is where most of the value is.
When should I automate something instead of running it myself?
Automate frequency, not importance. Something you run twice a month costs four seconds a month to start and becomes a system you have to maintain if you automate it. Prove the procedure by hand for a few weeks first, because that is where the awkward cases appear.
How do I tell if an AI product is actually doing anything?
Ask what exactly it runs and whether you can see the instructions, what starts it and whether you can stop it, and what happens when it fails and how you would find out. Most marketing in this category does not survive those three questions.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.