A prompt is one request. A skill is a method your assistant loads whenever a matching task shows up, so it survives past the current chat. An agent takes multiple steps toward a goal, decides its own next step, and uses tools to get there. The industry calls all three "agents", which is why buying the wrong one is so easy. Below is a test that settles it in thirty seconds.
Updated 15 September 2026. Catalogue counts and prices read from the live store on the day of writing; the product page is always the authority if they ever drift apart.
The thirty-second test for a fake agent
Before any of the definitions, here is the thing worth knowing, because it is the question underneath almost every "is this really an agent" argument.
That is the whole test. A scheduled summary email gets called an agent. A three-step automation gets called an agent. A great deal of what is currently sold as an agent is a fixed sequence dressed in autonomous language, and the test above separates them in one try.
The three, side by side
| Prompt | Skill | Agent | |
|---|---|---|---|
| What it is | One request | A method that loads on matching tasks | A process that decides its own next step |
| Lifespan | The current message | Every session, once installed | The run, until the goal is met |
| Who drives | You, every time | You, with the standards pre-loaded | It asks, you answer |
| Right when | You can state the whole job in one sentence | The job recurs and you want the same standards | Step two depends on what step one found |
| Fails at | Drift across a long chat | Silently not loading when phrasing misses | Compounding its own early mistakes |
| Price here | From $9 | $19 | $39 |
What is a prompt, exactly?
A prompt is one request. You paste it, you get an answer, the transaction is over. A good one carries the role you want the model to play, the context, the task, the output format and the constraints - but it is still a single turn of a conversation, and when you close the tab it is gone.
That is not a weakness. It is the right shape for work with a fixed scope: "write an employee engagement survey for a 50-person company focused on management effectiveness." One request, one deliverable, done. What you are paying for in a prompt pack is that someone already worked out the phrasing, the constraints and the failure modes so you do not rediscover them on your own time.
Prompts stop being the right tool the moment you find yourself pasting the same block of text for the third time this week.
Fixed scope, single deliverable - exactly what a prompt is for. Fill in the prospect, paste, get the email back in one turn. No install, no setup.
View the prompt pack →What is a skill, and how is it not just a long prompt?
A skill is a method the assistant loads when the task matches. That sentence is doing two pieces of work.
"A method" - not an answer and not a request, but how a particular kind of job gets done. What questions to ask first, what order to work in, what to check before shipping, what "good" looks like for that specific role.
"Loads when the task matches" - this is the part that separates it from a prompt. You install it once. From then on, when you ask about editorial calendars, the content-marketing method is already loaded. You do not paste anything, you do not remember to activate it, and it is still there next Tuesday. A prompt lives in one message; a skill survives across sessions.
Skills are the right pick when the work is a domain rather than a deliverable - open-ended, recurring, and easier to describe in your own words than to specify up front.
Content marketing spans strategy, calendars, briefs and repurposing - too broad for one prompt. Install Lena once and the method is loaded every time the subject comes up, in this chat and every chat after it.
View Lena →What is an agent, and how do you spot a fake one?
An agent takes multiple steps toward a goal, decides what to do next based on what it just found, and uses tools - search, files, connected apps - to get there. All three parts have to be present. Multiple steps alone is a workflow. Tools alone is a plugin. The deciding is the load-bearing bit, which is why the test at the top of this page is the one that matters.
Our agents are built to lead the process rather than execute a script: they run an intake, ask what they need one question at a time, then produce a deliverable calibrated to your answers. If you want to compare the platforms that host this kind of work - Claude, ChatGPT Agent, Gemini, Comet, n8n, Cursor - that is the best AI agents in 2026 roundup.
Where the $9 prompt writes one email, George asks about your ICP, differentiation and the objection you keep hitting - and what it asks next depends on what you answered - then builds the full multi-touch sequence.
View George →What does one task look like as all three?
Cold email, same job, three shapes.
- As a prompt ($9): a tested structure - hook, one-line value proposition, low-friction ask. Fill in your product and prospect, paste, get an email in under a minute. Right when you need one email for one audience.
- As a skill ($19): the outbound copywriting method stays loaded. Ask for a subject line on Monday and a rewrite on Thursday and you get the same standards both times, without re-explaining. Zoe is the cold-email version.
- As an agent ($39): an intake first - ICP, differentiation, likely objection, sequence length - then a complete multi-touch sequence with subject-line variants, follow-up logic and testing notes.
Same underlying model. Three genuinely different products.
Which one does your task need?
Four questions, and the first "yes" wins.
- Can you describe the whole requirement in one sentence? Prompt.
- Will you want the same standards applied again next week, without pasting anything? Skill.
- Would a competent specialist need to ask you questions before starting? Agent.
- Does the right second step depend on what the first step turns up? Agent, and nothing else will do.
Where does each one let you down?
Prompts fail on drift. Nothing carries between messages, so the tenth output does not match the first, and the fix - pasting more context every time - is exactly the tax you bought the prompt to avoid.
Skills fail on silence. A skill loads when the task matches, so when your phrasing does not match, it quietly does not load and you get generic output without being told why. And a skill will not go do anything: it improves how the assistant works, it does not take steps on its own.
Agents fail in three specific, expensive ways:
- They compound their own errors. A wrong step three becomes the foundation for steps four through eleven, and by the time you read it the mistake is load-bearing and stated confidently.
- They are expensive per task. Planning, searching, retrying and verifying burns far more tokens than a single answer, and vendors meter agentic runs separately.
- They are hardest to trust exactly where the stakes are highest. Anything irreversible - sending, paying, deleting, publishing - is where you most want the automation and least want the autonomy. Let the agent do the work and prepare the action. Keep the final click.
What do they cost, and what is in the catalogue?
2,000+ skills at $19. 300+ prompt packs at $9. 100+ agents at $39. Every one carries a 30-day money-back guarantee, so a wrong guess between the three categories costs you nothing but the time.
The price gap tracks depth, not file size: a prompt is one well-built request, a skill is a working method that persists, an agent is a process that runs an intake and adapts to the answers. Prompt packs span the range because a single-purpose generator and a full role library are not the same product.
How do you install one?
A skill is a ZIP file with SKILL.md at the root. In the Claude desktop app, open Customize → Skills, upload the ZIP, and toggle it on. After that it loads by itself whenever a matching task comes up.
Two things gate that menu. The one that catches most people is code execution, which you switch on as Code execution and file creation under Settings → Capabilities. Plan availability is the other, and Anthropic's own pages disagree: its help centre currently lists Skills on Free, Pro, Max, Team and Enterprise, while its Academy tutorial lists Pro, Max, Team and Enterprise - so if you are on the free plan, check Settings → Capabilities for your own account rather than taking either page's word for it. On Team and Enterprise an owner can disable it across the whole organisation, which is why the menu is sometimes greyed out on a work account - that one is an admin, not a bigger plan.
Where there is no Skills menu, or in ChatGPT or Gemini, open the ZIP and paste the contents of SKILL.md into custom instructions. You lose the automatic loading; the method still works. Prompt packs need no installation at all - paste and go.
In summary
One request with a fixed scope goes to a prompt pack like the AI Sales Email Generator ($9). A recurring domain where you want the same standards every time goes to a skill like Lena ($19). A job where the next step depends on the last one goes to an agent like George ($39). And if a vendor will not tell you which of the three they are selling, run the test at the top of this page before you pay. All run in Claude, ChatGPT, Gemini or Copilot, with a 30-day money-back guarantee.
Related Guides
100+ AI agents, for the work that adapts
When the right next step depends on what the last one found, only an agent will do. 100+ agents at $39 across sales, marketing, finance, recruiting, SEO, coding and ops - or start with a prompt pack at $9 or a skill at $19.
Browse all AI agents →KissMySkills is a marketplace of 2,000+ AI skills, 300+ prompt packs, 100+ agents & free tools for Claude, ChatGPT & any AI chat.
