Tonight you do the same, and I will walk you through every step, every click included. By the end of this evening you will have three things that did not exist this morning:
- 1An agent configured around you, not a generic chatbot
- 2The first real task it completed for you
- 3A setup that keeps it working week after week, without rebuilding anything
Your week ahead
Three priorities
- Quarterly report, due Thursday. Draft by Wednesday night.
- Three runs: Monday, Friday and Sunday mornings.
- Book the December family trip. Prices go up on the 1st.
One risk
- Thursday is full from 9 to 5. Anything left on the report that day will not get done.
One decision
- Phone and internet: switch. It saves about $28 a month, and the contract ends this week.
One draft, ready to send
- "Hi Anna, the report will be with you on Wednesday evening. Thanks, N."
Take your agent with you
The whole guide as one PDF: every prompt, the plan for the evening and the checklists. Leave your email and it downloads straight away, and the link waits in your inbox for when the evening is yours.
What you need tonight
- One paid subscription. This guide is built around Claude Pro. Claude Code comes with the same plan. Already paying for ChatGPT? Codex works too.
- A phone or a browser for steps 1 to 3 if you use the Claude app. Start on the couch if you like.
- A computer, Mac or Windows, for steps 4 and 5, which is where your assistant becomes an agent. On the Claude Code and Codex paths you use it all evening.
- About 60 minutes, in one sitting or over two evenings. This page remembers which tool you picked.
- Nothing technical. No code, no automation platforms, no stack of five subscriptions. Every click you might not have made before is spelled out.
Which tool are you using tonight?
Start here if you have never opened a terminal. Everything happens in the Claude app: your phone first, your computer for the last step.
- Plan: Claude Pro, $20 a month, or $17 a month paid yearly. The free plan cannot run scheduled tasks, and step 5 is a scheduled task.
- Install: the Claude app on your phone (App Store or Google Play) and the Claude desktop app for Mac or Windows from claude.com/download.
- Your two files are saved by hand: you copy them out of the chat. Every click for that is below.
Every step has a prompt you copy with one tap, then a box telling you what you should see in the reply, so you always know whether you are on track. Where the tools differ, the step shows a small switch with the version for yours.
Step 0
What you are actually doing tonight
About 5 minutes: two of reading, the rest setting up
You are not learning a new profession. You are hiring.
That reframe changed everything for me. I spent twenty years in project finance before I built my first agent, and the skill that mattered was not coding. It was the same skill I used hiring people. Explain who you are, hand over a task, give feedback, write down what works.
Before we start, one definition, because the words matter and most content online blurs them.
An assistant answers when you ask.An agent starts the work on its own.
Why a personal assistant, and two words you will meet tonight
The difference is not intelligence, it is initiative: who opens the conversation, you or it. Tonight has two acts built around that line. In steps 1 to 4 we build the core, a role and a memory held in two files you will own. In step 5 we flip the switch that makes it run on its own schedule. Core first, initiative second. That is the whole method.
The role we are building is a personal assistant. Not an app with that name, the real thing. It holds your week, prepares your briefing, queues the decisions you keep postponing, and drafts what needs to be sent. Everyone wants a personal assistant. Almost no one has one, because a good one costs a salary. Yours costs you this evening. You do not need a business, a niche or an audience to feel the result tonight. You only need a week, and you have one.
Two words you will meet tonight
- Prompt: the message you send. Every prompt on this page has a Copy button; tap it, then paste into the message box with a long press (phone) or Ctrl+V (Windows) or Cmd+V (Mac).
- A .md file: a plain text file. The .md ending (Markdown) only means it can hold headings and lists. Your agent will be made of two of them, me.md and assistant.md. Any notes app can hold the text, and any computer opens the file.
- Check your plan. Open Claude, go to Settings and look at your plan. It should say Pro or Max. If it says Free, upgrade there first: step 5 needs a paid plan.
- Phone: install the Claude app from the App Store or Google Play and sign in. This is where steps 1 to 3 happen.
- Computer: install the Claude desktop app from claude.com/download and sign in with the same account. You need it only for step 5, but doing it now means no waiting later.
- A place for your files: make a note called Personal assistant in your notes app. Tonight you paste two documents into it.
Before you move on
- Your plan is a paid one.
- The app is installed and you are signed in.
- You know where tonight's files will live: a note, or the
assistantfolder.
Step 1
The interview
About 20 minutes, the longest step of the evening
Your agent is only as good as what it knows about you. Instead of writing a memo about yourself, let Claude interview you. It is faster, and it asks better questions than you would think to answer.
Your assistantLet's start with your week. What are the fixed points, the things that happen every week whatever else is going on?
YouWork 9 to 6. Team call Mondays at 10. School run Tuesdays and Thursdays, so nothing before 9 those days. Gym Wednesday and Saturday mornings.
Your assistantGot it. Of those, which ones are truly protected, and which would you move for something important?
Open the Claude app on your phone and start a new chat (the pencil or + icon). Paste the prompt into the message box and send it.
I want you to interview me so you can work as my personal assistant. Ask me questions one at a time, wait for my answer before the next question. Cover: my typical week and fixed commitments, my priorities right now, the decisions I have been postponing, the people I coordinate with most, how I like my briefings structured, and my communication style in messages and emails. Keep it to about 12 questions. When we are done, compile everything into a single well organized document titled me.md that a future assistant could read to understand me quickly. Do not include anything I have not said.
me.md
My week
- Work 9 to 6, Monday to Friday. Team call Mondays at 10.
- School run Tuesdays and Thursdays: nothing before 9.
- Gym Wednesday and Saturday mornings. Protected.
Priorities right now
- Finish the quarterly report by the 24th.
- Back to running, three times a week.
- Book the December family trip.
Decisions I keep postponing
- Switch phone and internet provider, or stay.
- Which course to take this autumn.
- Hire a cleaner, or keep doing it myself.
People
- Anna, my manager: short messages, numbers first.
- Mark, my partner: shared calendar, prefers voice notes.
Briefings
- Top three risks first, then the week day by day.
- Bullet points, no long paragraphs.
My voice
- Friendly, direct, short. I sign off "Thanks, N."
Under Claude's reply there is a copy icon. Tap it, open your Personal assistant note and paste. If Claude shows me.md as a file card instead, open it and use its copy or download button.
If the interview feels like the longest part, it is. It is also the part you never repeat. Everything after this step is minutes, not effort.
Before you move on
- me.md is saved somewhere you can find it tomorrow.
- Keep this chat open. Steps 2 to 4 happen in the same conversation.
Step 2
The first task
About 10 minutes, and this is the fun one
Now the moment that made me a believer. Your agent reads what it knows about you and does its first real job.
In the same chat, paste this:
You are now my personal assistant. Using everything from me.md above, deliver three things. First, my weekly briefing: the coming week day by day, built around my fixed commitments, with my current priorities placed where I actually have capacity for them, and anything at risk flagged at the top. Second, my decision queue: the three decisions I am postponing, each with the realistic options, what each option costs me, and your recommendation with reasoning. Third, drafts of the two or three messages or emails I am most likely avoiding right now, in my voice, ready to send. Ask me at most two clarifying questions before you start, then deliver all three.
Do not judge the quality yet. First versions are drafts. Judging comes next, and it is a step, not a failure.
Step 3
The tune-up
About 10 minutes, two rounds of honest feedback
Here is what most guides skip. The agent will get things wrong. Mine scheduled deep work at 9 pm because I once said I answer emails late. Wrong conclusion, honest mistake.
This is not a bug in the method. This is the method. You would give a new hire feedback in their first week. Give it to your agent. Paste this, then speak plainly:
Some of this is off. I will tell you what is wrong and why. After each correction, update both the work and your understanding of me, and tell me in one line what you changed in me.md so it does not happen again.
Then list what bothered you, in normal human language. Too optimistic a schedule. Messages too formal. You recommended the safe option on a decision where I told you I want speed.
When the plan looks like something you would actually follow, send: Show me the final updated me.md in full. Copy it and paste it over the old text in your note. The old version is now out of date, so do not keep both.
Step 4
The agent writes its own manual
Five minutes. Yes, really
This step feels like a trick, and it is the strongest one in the whole evening. Your agent just spent the evening learning how to work with you. Make it write that down.
me.md
My week
- Work 9 to 6. Team call Mondays at 10.
- Nothing before 9 on Tuesdays and Thursdays.
Priorities right now
- Quarterly report by the 24th.
- Back to running, three times a week.
My voice
- Friendly, direct, short. "Thanks, N."
assistant.md
My role
- Personal assistant. I hold the week and prepare what needs deciding and sending.
Every Sunday I deliver
- The weekly briefing, risks first.
- The decision queue, with my recommendation.
- Drafts of the messages being avoided.
Corrections from the first evening
- Mornings are for focused work, not meetings.
- No deep work after 8 pm.
I start each session by asking
- What changed since last week?
Paste this:
Based on everything in this conversation, write your own working instructions as my personal assistant. Title it assistant.md. Include: your role in two sentences, the exact three deliverables you produce each week, my formatting and tone preferences, the mistakes you made tonight and the corrections, and the questions you should ask me at the start of each session. Write it so that a fresh copy of you could read me.md plus this file and perform at tonight's level immediately.
Copy assistant.md into your note under me.md, just like in step 1.
Then make them real files, because step 5 needs files, not text. Still in the same chat, send: Give me me.md and assistant.md as two downloadable files. Download both, ideally on your computer (open the same chat in the desktop app: every chat is there too). On a phone they land in Downloads or Files.
Step 5
Flip the switch
About 10 minutes
Until now you started every conversation. This step hands over the initiative, and that is when your assistant becomes an agent.
Personal assistant: weekly session
Every Sunday, 6:00 pm
Reads me.md and assistant.md, then prepares the briefing, the decision queue and the drafts.
Next run: this Sunday
- Update the Claude desktop app. Scheduled tasks now run in the cloud, on time, even with your laptop closed. Older versions ran them on your computer, so update first.
- Make the project. In the left panel open Projects, click + and choose Start from scratch. Name it
Personal assistant, add your two files, me.md and assistant.md, and click Create. Do not pick Use an existing folder: that keeps the project on your computer, and then the task runs only while the computer is on. - Start a new conversation inside the project and paste the prompt below.
- Confirm the schedule. Since September Claude has one message box for everything: send the prompt and Claude answers with a summary of the task and a Schedule button. If your app still shows a Chat / Cowork switch, choose Cowork first, or open Scheduled in the left sidebar, click New task, then Create with Claude.
Every Sunday at 6 pm, act as my personal assistant. Read me.md and assistant.md in this project first. Then deliver the full weekly session: my weekly briefing built around my fixed commitments with risks flagged at the top, my decision queue with options, costs and your recommendations, and drafts of the messages I am most likely avoiding. Leave everything ready for me to review in one place.
The two-minute update, once a week
Your agent knows what you told it tonight. It does not know what you decided on Tuesday. So once a week, before the Sunday run, tell it what changed. Paste this and fill in the gaps:
Before you plan next week, here is what changed. Decided: ... New this week: ... No longer a priority: ... Update me.md with this, then show me the new version in full.
Send Give me me.md as a downloadable file and download it. In your Personal assistant project, remove the old me.md and add the new one, so the Sunday run reads this week's version. About a minute.
Before you move on
- The task appears under Scheduled with Sunday, 6 pm next to it.
- Your two files are inside the project (or the folder) the task reads.
- You know where to look on Sunday evening.
What you built tonight
Count it. An agent that knows you, a completed first task, two files you own, and a schedule it keeps without you. That is a real system, built in one evening, with one tool. And you can now name the line most people blur. An assistant answers, an agent initiates. Yours initiates.
Now look at what you actually created, because it is bigger than one role. Tonight's weekly session is one scheduled task. But every function your assistant performs can become its own scheduled task: the briefing on Sunday, the decision queue on Wednesday, an inbox of drafts on Friday morning. In a full agent environment these are your sub-agents. Each runs on its own schedule, and all of them read the same two files. That is the real architecture. me.md and assistant.md are not notes, they are the management layer of your future staff.
Every task reads these two first
- Mon
- Tue
- WedDecision queuenoon
- Thu
- FriFriday drafts8 am
- Sat
- SunSunday briefing6 pm
And a staff needs someone to run it. One role above the tasks, the one that watches the whole week and decides what gets escalated to you and what gets handled quietly. You could call it your chief of staff. Your hire, your call on the title.
Now say it out loud
Give yourself the credit. The first step is always the scary one, and you took it on a weeknight, between dinner and sleep. Almost everyone is talking about agents right now. Almost nobody has built one. Tonight you crossed that line, and that is worth telling your colleagues about.
Three small rituals make it real. They take ten minutes, and they are the part people remember.
Give it a name
A named agent stops being a tool and becomes a hire, and you will notice the difference in how you both work. In the same chat, paste this:
You have worked with me all evening. Suggest five names for yourself that fit the way you work with me, each with one line on why. Then wait for my choice. From now on, sign every weekly session with that name, and show me the line to add to assistant.md so a fresh copy of you keeps it.
Give it a face
Now describe yourself as a character for a portrait: how you look, what you wear, where you sit, the light and the mood. Base it on how you work with me, not on robots from films. Write it as one ready prompt for an image generator, under 80 words, vertical 9:16 format, no text in the picture.
Introduce it in your stories
Put the portrait in an Instagram story and add this text. Change the name and the line about what it does for you.
Meet my agent, [name]. I built it tonight, in one evening, without a line of code. It plans my week, weighs the decisions I keep postponing and drafts the messages I avoid. From Sunday it does this on its own. Want yours? The free guide is at natalyarojkova.com/guide/start
One honest line about where this road goes. The agent you built tonight lives inside your AI app, and for the weekly rhythm that is exactly where it should live. Molen went further, and the difference is not where it runs, it is who speaks first. Molen lives in my messenger. He writes to me before I ask, answers when I text him from anywhere like a person would, reacts to events and not just to a calendar, and holds one continuous memory across everything he does for me. A scheduled task delivers and waits to be opened. A resident agent is present. Moving an agent in with you like that takes more than an evening. Tonight you built the part every resident agent starts from.
If something goes sideways
Claude asked all twelve questions at once.
Reply: One question at a time, please. Start again from the first one. It happens now and then, and one reminder is usually enough.
I closed the chat halfway through.
Nothing is lost. Every chat is saved: open the menu (or the left panel on a computer) and find it under your recent chats, on any device. On the Claude Code and Codex paths, whatever was already saved is sitting in your assistant folder.
I cannot find a download button for me.md.
Ask: Show me the full text of me.md in your reply. Copy it with the copy icon under the reply. To make a file on a computer: on Windows, paste into Notepad, then File, Save as, set Save as type to All files and name it me.md. On a Mac, open TextEdit, choose Format, Make Plain Text, paste, and save as me.md.
There is no Schedule button.
- Check the plan: scheduled tasks need Pro or Max, not Free.
- Update the desktop app and restart it.
- Older layout: open Scheduled in the left sidebar, then New task, and paste the prompt there.
- Meanwhile the manual fallback works: send Sunday reset in the project.
The task ran but says it cannot find my files.
Check that both files are inside the project the task belongs to. If you built the project from an existing folder, it only runs while your computer is on. The bulletproof fix: open the task, and paste the full text of me.md and assistant.md under its instructions.
Claude Code or Codex keeps asking for permission.
That is by design. Allow it to create and edit files in your assistant folder. If it asks to do something you do not understand, decline and ask: What would this do, in plain words?
Is it safe to tell an AI all this?
Share what you would tell a new assistant on their first day, and nothing you would not. No passwords, card numbers or other people's private details. Claude's memory has its own switch in Settings, Memory, and you can delete any chat.
Can I do this in another language?
Yes. Paste the prompts as they are and answer in your own language, or add one line: Answer me and write the files in German (or any language you like).
Keep your agent and this guide
Every prompt from tonight, the me.md example and the checklists, in one PDF. Save it next to your two files, and the next agent starts from here.
And keep the conversation open. Write to me and tell me two things: what did your assistant get wrong that step 3 could not fix, and what would you delegate to it next if you could? Reply to the email or message me on Instagram, whichever is closer. I read every answer.
I fired excuses. I hired AI. Tonight, so did you.
Natalya