Mission Log

Build a Personal AI Assistant with Markdown Files and Zero Code

A starter kit of markdown files that turns Claude into a personalized work assistant -- no coding, no setup, just conversation.

AI

Most people hear “AI assistant” and picture something they need a developer to build. A chatbot with a backend, an API integration, a custom interface.

But the most useful AI assistant I have seen recently is just a folder of text files. No app. No database. No code. A set of plain markdown files that, when loaded into Claude, turn it into a personalized work assistant that knows who you are, what you do, and how you like to communicate — and gets better every time you use it.

I have been calling it the My Assistant Kit.

TL;DR

  • A folder of markdown files that turns Claude into a personalized work assistant — no coding required.
  • CLAUDE.md acts as a hidden instruction file that tells Claude how to behave, what to ask, and how to learn about the user over time.
  • TODO placeholders in each file let Claude track what it knows and what it still needs to ask, faking persistent memory across sessions.

The Problem This Solves

AI tools are powerful, but they have an onboarding problem. The gap between “I signed up for Claude” and “Claude is actually useful to me every day” is enormous for most people. They open a chat window, stare at it, type something generic, get a generic response, and close the tab. The tool is not the bottleneck — the blank page is.

The people who get the most out of AI assistants are the ones who have learned to give them context. They write detailed system prompts. They structure their requests. They know how to shape a conversation so the model has enough information to be genuinely helpful.

But that knowledge is gatekept by familiarity with how language models work. The people who would benefit most from a personalized AI assistant — busy professionals drowning in tasks they cannot offload — are exactly the ones who do not have time to learn prompt engineering.

The My Assistant Kit sidesteps the whole problem. Instead of teaching someone how to prompt an AI, you give them a folder of files that do the prompting for them.

How It Works

The kit is a folder with this structure:

My-Assistant/
├── CLAUDE.md              ← The instruction brain (Claude reads this first)
├── START-HERE.md          ← Warm welcome + what to do first
├── ABOUT-ME.md            ← Who the user is and what they do
├── MY-WORK.md             ← What their job looks like day to day
├── MY-GOALS.md            ← What they want help with
├── MY-STYLE.md            ← How they like to communicate
├── TASKS.md               ← Recurring things they need help with
├── WORKFLOWS/
│   ├── daily-check-in.md  ← Morning routine template
│   ├── weekly-review.md   ← End-of-week reflection
│   └── project-helper.md  ← Project scoping and tracking
└── OUTPUTS/
    └── notes-and-summaries.md  ← Where Claude saves session takeaways

The user opens START-HERE.md, which tells them to open ABOUT-ME.md and answer a few questions. That is the entire onboarding. Claude takes it from there.

The CLAUDE.md Trick

The most interesting file in the kit is CLAUDE.md, and the user never reads it directly.

CLAUDE.md is a system prompt disguised as a document. When a user loads the kit’s folder as a Claude project, Claude reads CLAUDE.md automatically and immediately knows how to behave. It is not a configuration file — it is a set of behavioral instructions written in natural language that Claude follows because that is what language models do when you give them instructions in context.

Here is what CLAUDE.md tells Claude to do:

Greet the user by name once ABOUT-ME.md has been filled out, and reference something from their context to show it remembers them.

Check for incomplete files. Look for questions that are still marked TODO and gently suggest the user fill them in, one at a time, before diving into tasks. Never overwhelm — ask one question at a time and wait for the answer.

Update files as it learns. After any conversation where the user shares new information about themselves, their job, or their preferences, update the relevant .md file with what was learned. Always tell the user what was updated and why.

Treat every file as living memory. Reference and update the markdown files rather than relying on chat history. The files are the persistent layer.

Stay warm and non-technical. Celebrate small wins. Never make the user feel behind. If someone seems confused, slow down and re-explain using a simple analogy. Default to asking “what would be most helpful right now?” when unsure what the user needs.

The result is that the user’s first conversation with Claude feels like talking to an assistant who already knows how to help them — because CLAUDE.md has already done the instructing.

This is the same principle behind good software design — the user interacts with the interface, not the implementation. Except here, the “interface” is a warm welcome letter and the “implementation” is a markdown file that instructs an AI.

Faking Persistent Memory with TODO Placeholders

Language models do not remember between sessions. Every new conversation starts from zero. That is the fundamental limitation this kit has to work around.

The solution is simple: use the files themselves as memory.

Each questionnaire file (ABOUT-ME.md, MY-WORK.md, etc.) has placeholder text under every question:

**What is your first name?**

[TODO: Your answer here]

When Claude reads a file, it can see which questions have been answered and which still say TODO. That single pattern gives Claude the ability to:

  1. Know what it has already learned about the user
  2. Know what it still needs to ask
  3. Prioritize onboarding (filling in empty files) over task execution
  4. Track progress across multiple sessions

When the user answers a question in conversation, Claude updates the file — replacing the TODO with the actual answer. Next session, Claude reads the updated file and picks up where it left off.

It is not a database. It is not a vector store. It is a text file that a human and an AI both read and write. And it works — because language models are remarkably good at following simple conventions. The TODO pattern is a crude state machine built from placeholder text, and Claude treats it exactly like one.

The Self-Improving Loop

Every conversation makes the kit more useful. When the user answers a question, Claude updates the file. When they mention a new task, Claude adds it to TASKS.md. When they finish a weekly review, Claude updates MY-GOALS.md. The files accumulate context over time, and each session starts with more information than the last.

After a few weeks of use, the user’s ABOUT-ME.md and MY-WORK.md read like a detailed brief — because they are. The kit has progressively shaped itself into a document that tells Claude exactly how to be useful to this specific person.

The workflow templates in the WORKFLOWS/ folder reinforce this loop. A five-minute morning check-in, an end-of-week reflection, a project scoping template — each one gives the user a reason to come back, and each session feeds new context into the files.

Customizing for Any Audience

The kit’s structure is reusable. Only the questions change.

A version for small business owners might replace MY-WORK.md with something focused on clients and revenue. A version for students might swap in academic-specific workflows. You can also tighten or loosen the instructions in CLAUDE.md — make it more proactive, less verbose, more focused on a specific domain.

The best way to test a customized kit is to use it. Generate the files, start a new session with just CLAUDE.md and one questionnaire file, and see how Claude behaves. If it is too verbose, tighten the instructions. If it skips the TODO checks, make them more explicit. If the questionnaire feels stilted, rewrite the questions in a more conversational voice. The kit is just text — iterating on it is as simple as editing a file.

What This Points Toward

The My Assistant Kit is not a product. It is a pattern.

The pattern is: use markdown files as both the interface and the memory layer for an AI assistant, with a hidden instruction file that shapes the AI’s behavior without the user needing to understand how it works.

That pattern is more powerful than it looks. It means anyone can build a personalized AI assistant for any audience — their team, their family, their clients — without writing code. They just write the questions they want the AI to ask and the instructions they want it to follow, and the AI does the rest.

If you have been thinking about how to make AI useful for people who are not technical, this is a good place to start. Not with an app. Not with an API. With a folder of text files and a thoughtful set of instructions.

Getting Started

The kit is available on GitHub: subdepthtech/my-assistant-kit. Clone the repo (or download the ZIP), open the folder as a Claude project, and follow START-HERE.md. Claude will guide you through the rest — one question at a time, at whatever pace works for you.

What would you put in your own starter kit? What questions would you want an AI assistant to ask you before it started helping? Drop a comment — the beauty of this approach is that the answer is entirely up to you.