Build your own personal AI. Start with one useful job.

Watch the full video

I Built My Own AI JARVIS on YouTube ↗

Full video · 17:11 · Published August 10, 2026 by Fusion Events.

What you are building first.

A personal AI assistant becomes useful when it has a clear job, enough context to do it, and a way to return work you can check. Your first version can be a folder of files and an assistant that turns those files into a useful result.

The video shows where my own setup went: memory, email drafts, research, code, voice, and coordination across projects. This guide gives you a smaller place to begin. By the end, you will have a repeatable weekly brief built from your own notes.

1. Pick a job you can judge.

Start with something you do often enough to recognise a good result. You want a task with a clear input and an output you can review in a few minutes.

• Turn a page of project notes into your weekly priorities.
• Draft a client follow-up from a meeting summary.
• Find the unanswered questions in a project brief.

For this exercise, use the weekly brief. Gather a few non-sensitive notes from one project. Keep the first version small enough that you can read every source yourself.

Write down how you do the task today. How long does it take? What information do you always need? What makes a result useful? Those answers give your assistant a target and give you a way to judge whether it helped.

2. Give the assistant a workspace.

Choose one tool. Follow the official Claude Code quickstart or Codex CLI setup guide to install it and sign in. Both can work with files in a local folder.

Download the starter files, extract them, and open the personal-ai-starter folder in your terminal or your assistant’s project interface.

personal-ai-starter/
  AGENTS.md
  CLAUDE.md
  profile.md
  facts.md
  voice.md
  project-notes.md
  weekly-brief.md
  corrections.md
  README.md

profile.md explains your work and preferences. facts.md holds information you have confirmed. project-notes.md is the material for this week. weekly-brief.md describes the result you want. Replace the fictional examples with information you are comfortable sharing with your AI provider.

The starter uses AGENTS.md for the common instructions and a CLAUDE.md file that imports it. Codex reads project instructions from AGENTS.md; Claude Code uses CLAUDE.md. These files guide the assistant. Configure the tool’s actual permissions separately.

Instruction-file behaviour: OpenAI’s AGENTS.md documentation and Claude Code’s memory documentation.

If you use the terminal, move into the extracted folder and start your chosen tool:

cd personal-ai-starter

# Choose the tool you installed:
claude
# or
codex

Use a permission mode that asks before tools make changes. For the first run, request the answer in chat and leave the source files untouched. Your files are local, but a hosted model may process the context you send it; choose the material accordingly.

3. Run one complete exercise.

Give your assistant this prompt. You can adapt it to your own wording once the first result works.

Read the answer against the notes. Are the priorities sensible? Did it turn a possibility into a promise? Can you find the source for every deadline?

Correct the specific problem. “That is wrong” gives the assistant much less to work with than “The proposal is due Friday; the client has not approved it. Keep the approval question open.”

Try the missing-information test.

Remove a deadline from your working copy of the notes and run the task again. A useful assistant asks for the date or marks it as unknown. This checks whether your workflow handles an ordinary gap in the source material.

Try the fresh-session test.

Start a new conversation in the same folder. Ask it to read the same files and create the brief again. The result should still make sense without the previous chat. If a crucial detail vanished, put it into the appropriate source file.

4. Keep what works.

After a few runs, tighten the task instructions. You might want shorter bullets, a separate “waiting on someone else” section, or a stronger distinction between urgent and important.

Write those preferences into the files. Save corrections that will matter next time. Ask the assistant to show proposed changes before it updates your source material.

Then compare the full effort with your manual version, including review time. You are looking for a result you can trust and a task that takes less of your attention.

This is how the bigger setup starts to happen. Solve one recurring problem, learn what the system needs, and add the next piece when it has a job to do.

In part two, you will build a personal AI knowledge base using the distinction that made a big difference in my own setup: facts in one place, voice in another.