Create Your Own AI Assistant: Instructions, Tools, and Memory

Creating your own AI assistant is now a practical task for anyone with a recurring job they want help automating or accelerating. You do not need to train a model or write code. You configure an existing AI app with instructions that fit your work, give it the right access, and test whether it does the job reliably.

The part most people underestimate is persistence. A general-purpose chat session starts fresh every time. If you connect your assistant to a shared record like fenbs, it reads your current work context, projects, and rules before each session without you rebuilding them from scratch.

Decide what the assistant should do

Before opening any platform, write one sentence: "My assistant will ___." Fill in a specific, recurring job. Not "help with all my work." Something like "draft first versions of weekly status reports from my meeting notes" or "review pull request descriptions against our team's checklist and flag gaps."

A focused job makes the assistant testable. You know what a good output looks like. You know when it is wrong. You can improve the instructions based on specific failures. Vague instructions produce vague assistants that sometimes help and sometimes do not.

Choose a platform

Several AI platforms let you configure a custom assistant without building your own. Claude (Anthropic) offers Projects, which store shared instructions and files. ChatGPT offers custom GPTs with a system prompt and optional tool connections. Cursor and GitHub Copilot support configuration files for coding-specific instructions.

Choose based on the job. Writing and analysis work well in Claude and ChatGPT. Code tasks work well in Cursor, Copilot, or Claude Code. Research tasks work well in Perplexity or a model with search access. For the full range of options, AI personal assistant and AI employees cover tools across different categories.

Write clear instructions

Instructions are the most important part of creating a useful assistant. Write them as if briefing a capable person who is new to your work. Cover: what you are working on, the format you expect, the quality standards that matter, what to ask about rather than decide independently, and what to refuse.

Keep the instructions focused on the job you chose. Long, general instructions are harder to follow and produce inconsistent results. Add a section for limitations: topics the assistant must not address, actions it must not take without asking, and cases where it should say it does not know.

• The job: one sentence describing the specific task.
• The format: what a good output looks like.
• The standards: what matters in your work (tone, accuracy, completeness).
• What to ask: decisions and ambiguous cases that need your input.
• What to refuse: topics and actions outside the assistant's scope.

Add tools and data access

Many AI platforms support tool connections, letting the assistant read files, query databases, or call external services. Add only what the assistant needs for its specific job. An assistant that can read and write everything in your workspace is harder to control and more likely to cause an accidental problem.

MCP (Model Context Protocol) is a standard connection that lets an assistant read and update an outside tool. If your AI app supports it, you can use MCP to connect tools like task boards, knowledge bases, or code repositories. The assistant then reads current information from those tools rather than relying on static context you typed.

Test before trusting

Before you rely on the assistant for real work, run it through test cases. Use your most common inputs. Use edge cases that sit near the boundary of its instructions. Use cases the instructions do not cover and check whether it asks for clarification or guesses.

Expect to revise the instructions after the first round of testing. Most instruction sets need adjustment after you see how the assistant handles real inputs. That is expected. An assistant that handles your real test cases well is ready to help with real work.

Keep instructions current

Your work changes. Your assistant's instructions need to keep up. If a standard changes, if a new type of task comes in, or if you discover the assistant is handling a category of input poorly, update the instructions. If the instructions live only inside a chat app, they are easy to forget and hard to share with a colleague who needs the same assistant.

How to do this with fenbs

1. Create a free account. Add a project for the work your assistant will help with.
2. Write the rules your assistant should always follow as Decisions and rules. Include the format standards, the quality bar, and the things it must ask rather than decide.
3. Connect your AI app from Integrations. Claude, ChatGPT, Cursor, Codex CLI, Gemini CLI, GitHub Copilot, and Devin Desktop connect over MCP.
4. Ask the assistant to read the project context before its first task. Check that it understands the job correctly. Correct anything it misreads.
5. Before each session ends, ask the assistant to save Where we left off and update the relevant tasks. The next session starts from there.

fenbs does not configure your AI app. It holds the shared work context, rules, and progress your assistant needs. For platform-specific guides, JetBrains AI assistant covers one popular coding environment.