Which tasks to give the assistant
A personal AI assistant is an agent that works for one person: it knows your rules, documents and schedule. You give it repetitive work with text.
Good fits: email drafts from your templates, summaries of long documents, finding answers in internal rules and contracts, a daily digest of incoming email, and meeting preparation from the correspondence.
Poor fits: decisions you are responsible for, sending email without your review, payments and granting access. Here the assistant prepares a draft, and the decision stays with you.
The model runs in Ollama on your computer, under the MIT license. Drafts and documents do not go to an outside service.
If the assistant is not for you but for employees and clients, with work in a CRM and the 1C accounting system, it is a separate project: we build an AI assistant for business as a turnkey solution.
Step 1: role and instructions
First, install the platform and the model. Dify is a platform for AI apps and agents under the Apache 2.0 license with two conditions: without permission you may not run a multi-user cloud service like Dify or remove the logo from the console. These conditions do not affect a personal assistant.
1. Install Docker Desktop (Dify runs through it on macOS and Windows) and Ollama. Download the model with the command ollama pull qwen3:8b.
2. Download the Dify code from GitHub with the git clone command from the Dify documentation, go to the dify/docker folder, run cp .env.example .env and then docker compose up -d.
3. Open http://localhost/install, create an administrator account and sign in to Dify at http://localhost.
4. In “Integrations” > “Model Provider”, install the Ollama provider and add the model: qwen3:8b in “Model Name”, and in “Base URL” the Ollama address that the Dify container can reach. Your computer’s address on the local network works; localhost from inside the container does not.
5. Turn on “Function call support” (tool calling) for the model: without it the assistant cannot work with email and calendar.
6. In “Agents”, click “Create” > “Create from Blank”, name the assistant and open “Configure”.
7. In the “Prompt” field, write the system prompt: the instruction the model reads before each conversation. Describe the role, the order of work, when to use tools and the answer format, for example: “You are the assistant to the head of purchasing. Write briefly and politely. Prepare emails only as drafts. If unsure, ask.”
How to write such instructions is covered in detail on the page on the agent’s system prompt.
Step 2: your documents
Documents give the assistant knowledge of your work: internal rules, contracts, price lists, email templates. Dify offers two ways to add them.
1. Add a few reference files in the “Files” section inside “Configure”: the assistant reads them as it works.
2. For a large archive, create a knowledge base: “Knowledge” > “Create” > “Create a ready-to-use knowledge base”, then “Import from file”. You can upload up to 5 files at a time.
3. Set up how the text is split into chunks, choose the search method and wait until processing ends.
4. Search by meaning needs an embedding model: it turns text into vectors, sets of numbers that help find passages with similar meaning. Choose it in “Default Models” > “Embedding Model”.
Do not upload everything. An outdated rule gives an outdated answer, so delete old versions of files as soon as a new one appears.
Step 3: email and calendar
Email and calendar connect as tools: actions the agent calls by itself. In Dify, you add tools in the “Tools” section inside “Configure”: plugins from the Marketplace, your own APIs and MCP servers (a standard way to give AI access to a service).
The Email plugin from the Dify developers sends email over SMTP, with copies and attachments. It cannot read email.
The Outlook plugin reads, searches and sends email and works with the calendar: it shows, creates, changes and deletes events. To connect it, you need an app registration in the Azure Portal and a sign-in with a Microsoft account.
Outlook is a foreign service for a Russian user: emails and meetings pass through its servers. If your email is on a Russian server, n8n is more convenient: the “Email Trigger (IMAP)” node receives email from any server over IMAP, and the “Action” parameter set to “None” leaves them unread.
Start with reading and drafts. Turn on sending in your name after a week of checking drafts without finding errors.
What not to give the assistant
On 2 May 2023, Samsung banned generative AI for employees after engineers uploaded internal source code to ChatGPT. An assistant on your own computer removes this problem, but not all the risks.
Do not give the assistant passwords and access keys in the instructions or files, the right to send email and pay without your confirmation, or access to all company email when it needs only yours.
Correspondence with clients contains their personal data. Under 152-FZ, Russia’s personal data law, the data of Russian citizens is stored in databases in Russia, and before a transfer abroad the operator notifies Roskomnadzor, the Russian data regulator. So do not connect email with client data through foreign services without a lawyer; more on the page on data security.
Grant the minimum rights: read-only, only the folder it needs, and a person’s confirmation before any action. The full list of checks is on the page on agent protection.