How to build your own AI assistant and sell it to clients
The sign that it is time: you are writing the same long request for the tenth time. A prompt is a one-off instruction; an assistant is a permanent role you can hand to somebody else. Below is what one is made of and how it sells to a business.
The sign that it is time: you are writing the same long request for the tenth time. A prompt is a one-off instruction; an assistant is a permanent role you can hand to somebody else. Below is what one is made of and how it sells to a business.
---
Explaining the same thing every time
The sign that it is time: you are writing the same long request for the tenth time.
Every conversation starts with an explanation — who you are, who you work for, what format is needed, what to avoid. The model does not remember, so you repeat it. Half the time goes on entering context rather than on the task.
The second sign: it works for you and not for a colleague with the same tool. Because you know how to set a task and they do not. And explaining that in words does not work — it is easier to do it for them.
Both signs point to one thing: stop writing prompts and build an assistant.
How an assistant differs from a prompt
The difference is not cosmetic.
A prompt is a one-off instruction. You sent it, got an answer, and next time you write it again.
An assistant is a permanent role. It has who it is, what it knows, how it behaves, what questions it asks and in what form it returns a result. You arrive with a task in ordinary words — the context is already inside.
Three practical consequences.
No repeating. The role, the constraints and the format are set once.
It asks by itself. A good assistant does not try to guess what is missing — it clarifies. That is what separates it from a prompt, which silently fills gaps with guesses.
It can be handed over. To a colleague, an employee, a client. Passing on a prompt is useless — the person does not know what to change in it. An assistant gets used with no instructions.
What it is made of
Four parts. Skipping any one gives you "another chatbot".
The role. Who it is and what it knows about. Specifically: not "a marketer" but "a specialist in product listings for small brands".
The knowledge. What it has to know about you or your business: services, rules, examples, what is allowed and what is not. An assistant with no base answers in generalities — the main reason home-made helpers disappoint.
The behaviour. What it does first, what it clarifies, what it never does, how it behaves when data is missing.
The output format. What exactly it returns: a document, a list, a table, a draft. Without that you will get whatever the model finds convenient.
The technique that speeds everything up
Assembling an instruction by hand is slow, and the first version always comes out badly.
A meta-approach works: ask a model to write the instruction for the future assistant. You describe the task in one line and out comes a professionally formatted instruction with all four parts.
Then you adapt it. Correcting something finished is many times faster than writing from scratch, and the structure is right immediately — you will not forget the constraints and the format that everybody forgets from scratch.
The second technique is feedback from the model itself. You submit the finished instruction for review: what is vague here, what is missing, where the assistant will start improvising. Corrections from that list usually improve the result more than another hour of your own work.
Where the content comes from
The most underrated part: assembling an instruction is easy, filling it with knowledge is not.
Interviewing the person who holds it. If the assistant is for a company, the knowledge sits in employees' heads. It gets extracted with questions: what do you do, in what order, what goes wrong, how do you handle exceptions. The person does not have to write anything — they have to answer.
Existing material. Price lists, procedures, message threads, frequent questions. There is usually more of it than expected and it is scattered across different places.
Real client conversations. The best source of phrasings: how people actually ask. An assistant built on live questions answers recognisably.
And what must not be in the base: personal data, commercial information that cannot leave the building, and anything you are not prepared to answer for if the assistant repeats it to a client.
How this sells
Here the money side begins, and it is more interesting than personal convenience.
A business needs not a chat but a helper that knows their specifics. A general model does not know your services, your prices, your rules. An assistant built for a company does — and that is the whole value.
What gets bought most often: a helper answering customers that knows the services and prices; an assistant for a department with its procedures; a helper for staff answering from internal rules instead of them interrupting colleagues; a sorter for incoming enquiries.
Why people pay for it. Not for the technology — for somebody having assembled the company's knowledge into a working form. The main work here is not technical but extracting knowledge: interviews with people, gathering rules, structuring.
From my scrape of 848 listings: one-off work runs at a median of $30–75, while a system or pipeline is $490–1 460. An assistant for a business is the second category, and the difference is not the difficulty of building but that a result gets sold rather than an operation.
The limits
An assistant does not replace a process. If nobody in a company knows how anything gets done, an assistant will not fix that — it will cement the chaos.
It does not carry responsibility for decisions. Everything verifiable gets verified; final decisions with consequences get made by a person.
And it needs support. Prices or services changed — the base is out of date. An assistant nobody watches starts lying to customers within six months.
How to tell the assistant came out right
Four checks before it goes into work or to a client.
It asks questions rather than guessing. Give it an incomplete task. A good assistant will clarify what is missing; a bad one will fill the gaps with invention — and in real work that will cost dearly.
The result is stable. Run the same type of task three times. If the format differs each time, the instruction is vague.
Somebody else can use it. Give it to a colleague with no explanation. If they had to ask how to handle it, the assistant is unfinished.
It knows its boundaries. On a task outside its remit it should say that is not for it rather than attempt it. That gets written in separately and checked separately.
And last: build in updating. Prices, services and rules change. An assistant nobody watches starts confidently lying within six months — and worst of all if it is at a client's by then.
Where to start
The "Personal AI bureau" express course — the move from being a user to being somebody who designs assistants and sells them to clients. Inside is the "instruction generator" meta-prompt: you describe the task in one line and get a ready professional instruction for your own assistant, formatted properly.
The result: you assemble your assistant in a minute through the meta-prompt and understand the architecture of instructions from inside — meaning you can fix and extend them rather than only copy somebody else's.
The price is $9. The course is included in the Creator plan — not the base one, and it is not in the three free days of Basic.
Three routes to it: a one-off purchase at $9, the Creator+ plan, where it sits alongside building your own assistants on the platform and the advanced skills, or 1 000 experience points.
What to do first. Before building your own, look at the ready ones: more than 140 AI assistants across 14 categories — marketing, sales, HR, legal, e-commerce, analytics, engineering. A substantial share of the tasks people set out to build their own for is already covered. The chats come with the Basic plan, and registration is free and gives three days of Basic.
The rule is simple: you build your own when a ready one did not fit — not the other way round.
Do one thing today: find a request you have written more than three times. That is your first candidate for an assistant, and building it is half an hour's work.
---
The five lessons the skill consists of
Module 9, "AI Assistants": system instructions for GPTs · copywriting · conversation design · marketing · SendPulse.
The last is the most applied: it covers the sales funnel, building a sales bot that walks a client through a staged funnel, and a consultant bot for a website and Telegram. Plus packaging a bot for a business from scratch, from idea to finished product.
The first lesson draws on section 5.7 of Andriy Burkov's book — the key components of a system instruction.
A ready helper that writes instructions
The bot architect from the archive designs system prompts: the role, a seven-step reasoning chain and a block of hard prohibitions.
The "factory of 100+ bespoke assistants" mini-course comes with the Creator plan.
Module 9 — Creator; the architect — Make, N8N and Full.
---
What to read next
[Anatomy of a system prompt](/en/blog/system-prompt-anatomy) — the six blocks in detail.
[A company knowledge base as a service](/en/blog/company-knowledge-base) — the most labour-intensive part.
[What an assistant for a department costs](/en/blog/assistant-for-department-cost) — how to price it.
[One assistant against three](/en/blog/one-assistant-vs-three) — when you need a chain rather than one.
[AI agents: what they are and what people pay](/en/blog/ai-agents-guide) — the full guide to agents, with market rates.