The Tool Ladder: Prompt, Project, Skill, Agent

You have seen the words Skill and Agent enough times to feel that you should already know the difference, and not once has anyone explained it in a way that told you which one you need on Thursday.

If that is where you are, you are not the only one. The words get sold as levels of sophistication, as though an Agent is what you graduate to once you are good enough. That is the wrong picture, and it is why the choice feels harder than it is.

Here is the picture that works. These are not four levels of skill. They are four amounts of your own thinking, saved in four different places. I set it out as a ladder in Book 5 of my Claude book, and it is the frame I reach for whenever someone asks which one they need.

The four rungs.

• A prompt is one request. You describe the job in full, you get an answer, and nothing is kept. Everything you knew about the task stayed in your head and in that one message. Best for work you will do once.

• A Project is a room. A workspace that holds your standing instructions and your background material, so every conversation inside it starts with the situation already understood. Best for a category of work you come back to.

• A Skill is a saved method. Where a Project holds your context, a Skill holds your whole approach to one specific job. You call it, give it the input, and get the same shape of output every time. Best for one task that repeats in the same form.

• An Agent is a standing job. A helper that pursues a goal across several steps, inside rules you set once, with some independence. Best for a responsibility rather than a task.

Read down that list and notice what actually changes. It is not the cleverness. It is how much of the thinking has moved out of your memory and into something that holds it. That is why the ladder exists, and it is why the top rung is not the prize.

Two questions decide the rung.

You do not need to weigh features. Ask these instead.

How often does this job come back, in the same shape? Once is a prompt. Now and then, in a recognisable category, is a Project. Weekly, in almost identical form, is a Skill.

And how much of it do you want happening without you watching? If the answer is none, you are not looking for an Agent, whatever the job looks like. An Agent is what you reach for when you want something carried on while you are elsewhere.

Most work stops at the second question. Which brings me to the part the tool makers have no reason to tell you.

Higher is not better.

Using an Agent for a one-time task is like hiring a moving crew to carry a single box upstairs. It will work. You will also have spent an afternoon arranging it, and the box would have taken you ninety seconds.

Every rung above the first has a setup cost, paid once, that only makes sense if the job repeats often enough to pay it back. A Project earns its ten minutes within a fortnight if the work is genuinely recurring. A Skill earns its setup if the task is truly identical each time, and not merely similar. An Agent earns it when the alternative is you remembering to do something every day.

Honestly, most of what most people do belongs on the first two rungs, and there is nothing unambitious about that. The person with three well-written Projects is getting more out of this than the person with an Agent they no longer trust.

Free Claude toolkit. The companion pack for my Claude book is free: ready-made Project instruction templates, and more than five hundred numbered prompts you can copy and paste. No card, and you can leave the list in one click. Get the toolkit →

The rule that matters as you climb.

There is one thing worth knowing before you go past the second rung, and it is the opposite of what people expect.

The less you supervise, the sharper your standing rules have to be. When you are reading every reply, a fuzzy instruction is survivable, because you catch the drift as it happens. In something that runs on its own, that same fuzzy instruction repeats its fuzziness on every single run, unwatched.

So the writing gets harder as the automation gets easier. That is the real cost of the upper rungs, and it is not the setup screen.

What this will not do.

The ladder will not tell you whether a job is worth doing at all. It sorts jobs you have already decided to keep.

It will not rescue a task you cannot describe. If you cannot write down what good looks like, saving that description somewhere does not improve it, it just repeats it.

And it will not stay still. The names on these rungs belong to products, and products get renamed and rearranged. The four amounts of saved thinking underneath them do not change, which is why I would rather you remembered the ladder than the labels.

One thing to do before the next post.

Write down the three things you brought to Claude this week. Beside each one, put a rung: prompt, project, skill or agent.

Two of them will be prompts and should stay prompts. The third is usually the one quietly asking to become a Project, and that is ten minutes of work that pays back for months. If one of them turns out to be a small tool rather than a piece of writing, build it instead.

If you want the whole system rather than one piece of it, that is what I wrote Claude AI Bible (7 Books in 1) for. Seven books in one volume, with the full ladder, twenty-four Skill and Agent templates, the failure rules for anything you let run unattended, and more than five hundred numbered prompts to copy.

This post is part of my Claude series. The one that started it is Why Claude Gave You a Great Answer on Monday and a Flat One on Tuesday.

Tomasz