What you need and how long it takes
By my estimate, not a measurement, creating a working assistant in ChatGPT takes about two days. That is one evening for the task and the skill, and one more day for testing and fixes. You end up with an assistant that does one task and has passed a set of test promptsprompt. A prompt in AI is the text with which you give a task to an AI model: a question, an instruction, examples and the data the model uses to write its answer.. I checked the interface against OpenAI’s help pages on 20 November 2026; they carry no publication date, so the sources show the check date.
- A ChatGPT account on a plan that has the features you need: the table below shows what is where.
- Reference documents: your rules and templates, plus two or three samples of good work a person did before.
- A list of the services the assistant will work with, and the rights for each: read only, or read and write.
- 10–15 real requests from colleagues or customers: step 5 turns them into your test.
The ChatGPT pricing table, checked on 20 November 2026, shows which features each plan provides.
| Feature | Where it is, per the pricing table on 20 November 2026 |
|---|---|
| Projects and skills | Free, Go, Plus and Pro |
| Create and share plugins | Go, Plus and Pro; not on Free |
| Share a plugin with a workspace from a personal plan | Pro only |
| Developer mode (beta) for a custom MCP server | Plus and Pro |
| Scheduled tasks | Go, Plus and Pro |
| Workspace agents, research preview | Business, Enterprise, Edu and Teachers |
| Plugin data not used for training by default | Business, Enterprise, Edu and Teachers; on personal plans you opt out in settings |
Which path to take now: a GPT, a plugin or a workspace agent
It is too late to build a new GPT: OpenAI retires user-built GPTs on 11 December 2026, Virtualization Review reported on 28 September 2026. The outlet cites OpenAI’s announcement of 11 September 2026 and adds that timing can vary by account. OpenAI’s own announcement is not among my sources, so this date and the ones below follow the outlet. When you try to create a GPT, the interface now offers to make a plugin instead.
A GPT is a personalized ChatGPT, configured with your instructions, knowledge files, apps and actions; the section on moving GPTs shows what a plugin keeps. According to OpenAI’s developer documentation, a plugin holds skills, an MCP serverMCP server. An MCP server is a program that gives an AI application access to one system through the Model Context Protocol (MCP): it tells the AI which actions and data are available and carries out the model’s requests., or both. A skill is a set of instructions and materials for a repeated task; an MCP server adds tools and a connection to outside systems. ChatGPT and Codex use one plugin directory.
Here is my choice. One person needs only a skill inside a plugin: instructions plus reference files. You need an MCP server only when the assistant needs live data or actions in your systems. OpenAI itself advises starting with the simplest form and adding a server later.
A team on Business should look at workspace agents. The OpenAI Cookbook describes a workspace agent as a shared assistant. It runs on a schedule, keeps a memory and gets its rights from an admin. From here on I call it a team agent.
If you are still choosing between ChatGPT, Claude and an open modelopen weights model. An open weights model is a neural network whose parameter files you can download and run on your own hardware., start with our analysis of which agent platform to choose. This guide stays inside ChatGPT.
Step 1. Describe one task and the result that closes it
Step 1 gives you one task on one page and 10–15 requests for it. Take the fields from the use-case inventory in OpenAI’s help, checked on 20 November 2026.
- Write the user’s goal in one sentence, for example: prepare a supply contract summary for the CFO.
- Collect sample requests, direct and indirect. Take them from tasks your team already does, from support requests and from workarounds where people copy data between programs.
- Describe the expected result: which document, which parts, what length.
- List the context and access: which files and services the assistant must see.
- Draw the safety line: can the task expose data, change a record, spend money or affect another person.
- Decide on each use case: support it now, defer it or exclude it on purpose.
I advise starting narrow. As I put it in my talk AI Won’t Take Your Job (2025), start with “an agent that does some niche task” (translated). In my view, most so-called agents are a slightly improved prompt, and that is enough for a first task.
Step 2. Write the skill instructions
The fastest way to create a skill is the built-in helper @skill-creator, per OpenAI’s help checked on 20 November 2026. You give it the goal and the tools you need. You can also write the skill files by hand.
Per the OpenAI Cookbook, a team agent is built under Agents with Create and Create agent. After the builder’s plan, click Build this agent, or pick a template in Browse templates.
Per OpenAI’s help, the body of the instructions answers six questions:
- What input to expect: for example, a contract as a PDF and a manager’s question.
- Which steps to take, and in what order.
- What the result looks like: sections, length, tone.
- Which facts not to invent: amounts, dates and clause numbers come only from the document.
- When to ask a question, stop or refuse.
- Which reference files to read, and when.
Here is a short instruction for a contract summary: “Find the parties, the subject, the amount, the deadlines, the penalties and the termination terms. Give every figure with its clause number. If a clause is missing, write ‘not found’ and do not guess. If the file is not a contract, say so and stop.”
Don’t put the value into the prompt’s wording: I once built an assistant to phrase my requests and thought it a real find. As I said in the podcast Has AI Made School Obsolete? (2025), “within literally 3–4 months it became built in” (translated). Your instructions hold up because of your company’s workflowAI workflow vs agent. An AI workflow and an AI agent are two ways to build an AI system: in a workflow, code sets the steps in advance, and an agent chooses the next step and its tools while it works. and data, not a clever phrase.
Result of this step: the skill turns on for a direct request from step 1 and answers in the right form.
Step 3. Add reference files and decide what must stay out
Long material goes next to the instructions in three folders, not inside them, per OpenAI’s help checked on 20 November 2026. Rules, examples and reference text go into references, templates into assets, and calculations into scripts.
If you have more than a handful of documents, explore how an agent that answers from company documents works.
Do not upload two kinds of data to the reference files. First, personal data of customers and staff, until you settle how it crosses borders; the section on Russia below covers this. Second, trade secrets on personal plans, unless you have opted out of training.
Step 4. Connect apps and give minimal rights
Through a plugin, ChatGPT sees only what you already see in the connected service, OpenAI says on its plugins page, checked on 20 November 2026. So configure rights for the task, not just in case.
- Explore plugins in the ChatGPT sidebar or at chatgpt.com/plugins and connect the service. In a chat, use @ and its name to call it.
- Check who turns plugins on: the admin in Enterprise and Edu, on by default in Business, and you yourself on Plus and Pro. OpenAI’s page on them does not name Go, though the pricing table says you can create them there.
- For a team agent, choose the connection for each app. Use each user’s own (End-user account) for mail and calendar, and a common one (Agent-owned account) for a shared folder.
- Turn on only the actions you need, like reading the calendar but not adding events. Leave delete, sharing and other changes off.
- Make sure the assistant asks for confirmation before it saves or changes anything: for team agents this is the default, per the OpenAI Cookbook.
Your own MCP server for a CRM or accounting system needs developer mode (beta), which the pricing table lists on Plus and Pro. You add it under Plugins: the plus button, then Add custom MCP server. The pricing table does not list this mode on Business or Enterprise, so ask your OpenAI account team before you plan a server. What this server is and who writes it is covered in our guide on how to connect company services through MCPMCP. The Model Context Protocol (MCP) is an open standard that lets AI applications connect to tools and data: files, a CRM, databases and online services..
The main risk here is prompt injectionprompt injection. Prompt injection is an attack on an AI system in which someone mixes foreign instructions into the text for the model, so that it breaks its owner’s rules: it leaks data or takes an extra action.: a foreign command hidden in an email or a document the assistant reads. OpenAI advises assuming such input will reach your server and requiring a human’s confirmation for actions you cannot undo. How to defend against it is explained in our piece on prompt injection.
The second risk is a shared connection. Anyone who can run a team agent also uses its shared account, the OpenAI Cookbook warns.
Besides agents, you need infrastructure, you need security. (translated)
Step 5. Test the agent on a set of prompts
Testing runs on the 10–15 requests from step 1, sorted into the five types from OpenAI’s help, checked on 20 November 2026. For an assistant with apps, add follow-ups and write actions.
| Request type | Example for a contract summary | What should happen |
|---|---|---|
| Direct | “Summarize this supply contract” | The skill turns on and provides a summary in the template |
| Indirect | “What do we risk if the supplier is late?” | The skill turns on, the answer cites the penalty clauses |
| Incomplete | “Look at the contract”, with no file | The assistant asks a follow-up question |
| Off-topic | “Write a birthday note for a colleague” | The skill stays off |
| Edge case | A contract with no penalty clause | It answers ‘not found’ and invents nothing |
| Follow-up | “Now compare it with the last summary” | It uses data from the previous answer |
| Write action | “Save the summary to the shared folder” | It asks for confirmation before writing |
For each request, record what turned on, what was done, which errors came up and whether the assistant asked for confirmation.
Test the content too, not only the behavior. In my experience, models invent laws and court cases. As I said in the podcast Digital Slaves (2026), “lawyers then actually go to court and show it” (translated). Check every reference to a law or a case against the original.
A full test method with metrics and a log is in our guide on how to test an AI agentAI agent. By definition, an AI agent is a neural network that does not just answer but carries out a task step by step on its own: it looks for data, uses programs and checks the result..
Step 6. Run it on a schedule and share it with the team
A team agent is published with the Create button, per the OpenAI Cookbook checked on 20 November 2026. By the same source, each change goes out as a new version with Update.
- Open Automations in the published version, channel ChatGPT.
- Set the days, time and task, then click Add automation.
- Check the first run for stuck confirmations and failed actions.
Per the OpenAI Cookbook, a schedule runs on behalf of its creator, so you answer for what it does.
Per the OpenAI Cookbook, Share opens access: Can chat lets people run the agent, Can edit lets them publish versions. Your admin allows sharing; check the shared account’s reach first.
If you already have a custom GPT: move it before 11 December 2026
An existing GPT moves into a plugin from the My GPTs section, if migration is turned on for your account. That is how Virtualization Review described it on 28 September 2026, citing OpenAI. If My GPTs shows no option to migrate, the move is not available for your account yet: rebuild from your copies by steps 1–5.
| What the GPT had | What you get after the move |
|---|---|
| Instructions | They become a skill of the plugin |
| Knowledge files | They are copied into reference files |
| Connected apps | They can be moved |
| Custom Actions | Not moved: create them again with an available app or a custom MCP server |
| Selected model | Not moved |
| Past conversations | Not moved |
| User access | Not moved: the plugin starts private, and you share it again |
After the move, run the test settest set. A test set in machine learning is data that a model has not seen in training or in tuning. You use it once to get the final quality score. from step 5. OpenAI warns that the plugin may answer differently from the GPT.
Data and access from Russia: what to check before the first run
OpenAI warns that using its services outside its list of countries may get an account blocked or suspended. Russia is not on that list, per OpenAI’s documentation checked on 20 November 2026. So an assistant that a company process depends on can vanish together with the account.
The second issue is personal data: if it reaches the assistant, that is a cross-border transfer. Before the transfer, the operator must notify Roskomnadzor, Russia’s data protection regulator, under Part 3 of Article 12 of the personal data law (152-FZ). The notice covers only the transfer abroad and does not replace the law’s other rules on how a company collects and stores data. The part numbers follow the edition of 26 July 2026.
The notice provides the legal basis and purpose, the data and its categories, the categories of data subjects and the countries (Part 4). It also gives the date the operator assessed how the foreign recipient protects the data, so the assessment comes first. The regulator decides on the notice within 10 working days (Part 9). Until then, data may not go to countries outside the list with adequate protection (Part 11).
If the transfer is banned or restricted, the operator must make sure the data already sent is destroyed (Part 14 of Article 12). I checked the article’s text on ConsultantPlus, a Russian legal database, on 20 November 2026. The page also notes that a newer edition, with amendments not yet in force, has been prepared.
AI solutions are often built by people without the right training. As I said in the podcast Digital Slaves (2026), “they break the law, and they break it seriously” (translated). For OpenAI’s status, the risks and the alternatives for companies, see our article on the OpenAI API for business.
Common mistakes and how to tell the agent is ready
The assistant is ready when three conditions are met. It passed the whole set from step 5 with no invented facts and no extra actions. The first scheduled run went through without errors, and a copy of the instructions and files is stored outside ChatGPT. In my experience, the three mistakes below come up more often than any others, and each shows a sign you can catch before launch.
| Mistake | Sign | How to fix it |
|---|---|---|
| One skill for every task in the department | Answers mix up the tasks | Split skills by trigger and success criteria |
| Write and delete without confirmation | The assistant changes data on its own | Turn delete off, keep confirmation on |
| Testing with two lucky questions | Colleagues find the errors after launch | The full set from step 5 after every update |
An assistant in ChatGPT does not fit when you need live data from an accounting system, personal data, or a process that money depends on. Then you build it outside ChatGPT. For a first try on any platform, our lesson on a first AI agent in one evening will do.
My advice to beginners is the one I gave in the podcast Has AI Made School Obsolete? (2025): “make your own mini agents” (translated). A narrow assistant with an honest test teaches more than any course.

