How to Use AI as a Personal Assistant

• Tell it how you work: your role, the people you deal with, and how you want answers formatted.
• Connect the places your work lives: email, calendar, and documents.
• Give it what was said, too. Meetings, calls, and quick conversations are where most decisions and promises happen.
• Hand over one recurring job, with one sentence of purpose and a fixed format.
• Put it on a schedule. A brief that arrives on its own beats one you have to remember to ask for.
• Approve what leaves your hands. Let it draft, and you send.

Most of what a human assistant does comes down to briefs, drafts, and follow-through. Each job below ends with what the assistant needs to see to do it well.

• Morning brief: today's meetings, what you owe people, and what's waiting on others. Needs your calendar, email, and recent conversations.
• Meeting prep: where you left it with this person, and what you promised. Needs your past meetings with them.
• Inbox triage: what needs a reply today, with drafts in your voice. Needs your email.
• Follow-up after a meeting: a recap with decisions, owners, and dates. Needs the conversation itself.
• Commitment tracking: every promise you made this week, and to whom. Needs your conversations and email.
• Research: options compared against your constraints. Needs the web, plus what you said you need.
• Weekly review: what moved, what slipped, and what's next week. Needs all of the above.
• Personal admin: appointments, travel, and family logistics. Needs your calendar, plus what the doctor or the school actually said.

Look at what each job needs. Email and calendar cover the logistics. The jobs people value most, like prep, follow-ups, and remembering promises, depend on conversations. Those are the records most assistants never see.

ChatGPT, Claude, and Gemini all keep instructions or memory across chats, and most let you group work into projects. Put the things that don't change there, so you don't have to repeat them in every request:

About me: [Role] at [company]. I work mostly with [teams, clients, or people]. How I like answers: Short. Headings, then bullets. Lead with what needs a decision. Rules: Keep facts separate from suggestions. Don't invent owners or deadlines; if they aren't in my records, say so. Draft emails and messages for me to send. Don't send them yourself.

Keep it to what changes the answers. Your role and preferred format matter more than a long biography. Leave out facts that change week to week, like a deadline or a project's status. Those should come from your connected sources, so the assistant always works from the current version.

Most assistants now connect to email, calendar, and cloud documents. Connect one source at a time, starting with whatever you check most. Then ask, ‘What can you see in my calendar this week?’ to confirm the connection works before you build anything on it.

Start with read and draft access. Once you trust its drafts, you can decide how much more to hand over.

Your connected tools hold what someone typed. The decision you made at the end of a call, the ‘I'll send that by Friday’ on the way out of a meeting, and the instructions the doctor gave you never reach your inbox. A human assistant would have heard them. Your AI assistant needs a way to hear them too.

Pastok captures those conversations. Tap once on your Apple Watch or iPhone, and it records the meeting, a call on speaker, or you talking through a plan on a walk. Ask the people you're with before you start. Afterward, each conversation becomes part of a daily report:

• Activity: what happened, conversation by conversation.
• Action Board: every promise made out loud, with the date and the context it came from.
• Memory: the people, projects, and preferences Pastok has learned about you.

Next, connect Pastok to the assistant you already use. The Pastok MCP integration works with Claude, ChatGPT, Claude Code, Codex, and Cursor. Setup is one browser sign-in per tool. Your assistant can then search your recent transcripts by meaning, read your daily reports and Action Board, and use what Pastok remembers about you. For anything older, Pastok's own Chat searches your full history.

With Pastok connected next to your email and calendar, your assistant can answer questions no single tool can:

• ‘Check my Pastok Action Board and my inbox. Which promises from this week haven't I followed up on?’
• ‘Draft a follow-up email to [name] from today's meeting. Use the decisions and dates we agreed on.’
• ‘Compare [document] with this week's discussions about [project]. What has changed since it was written?’

Pick one job from the list above and write it as an instruction with a fixed format. A morning brief is a good first job, because it uses every source and you'll know by 9 a.m. whether it helped:

Prepare my morning brief from my calendar, my inbox, and my Pastok conversations from the last few days. Three sections: Today (each meeting, with one line on where I left things with that person), I owe (promises I made, with dates, and whether I've followed up), and Waiting on others. Under 200 words. Leave out anything you can't trace to a source.

A brief built from that instruction looks like this:

• Today: 10:00 product review with Sam. Last Thursday you agreed to cut CSV export from the first release.
• Today: 14:00 call with your accountant. She asked for the Q3 receipts.
• I owe: Q3 receipts to your accountant. You said Friday, and there's no email yet.
• I owe: an intro between Priya and Leo. You said ‘this week’ on Tuesday.
• Waiting on others: the contractor's revised quote. He said Monday.

Most of that brief came from things people said, not things they wrote down.

Other good first jobs:

• Prep for a recurring meeting: ‘Before each 1:1 with [name], tell me what we agreed last time, what I owe them, and what's still open.’
• A next-session brief for a project: ‘Tell me what changed since I last worked on [project], what's unresolved, and where to pick up.’
• A Sunday commitments list: ‘List every promise I made this week, to whom, and by when.’

When the output is wrong, correct the instruction rather than the answer. ‘Only list promises I made, not ones other people made’ fixes every future brief. Add a rule only when you've seen the problem it solves. A short instruction is easier to keep accurate than a long one.

A brief you have to remember to request is just another task. Once the format works, let it arrive on its own.

If the job needs your email and calendar, schedule it in your assistant. ChatGPT, Claude, and Gemini can all run a saved request at a set time.

If the job is about your conversations, save it as a Pastok Skill. A Skill is an instruction you write once in Pastok Chat, in plain language. It can run when you ask, on a schedule, or when Pastok detects a certain kind of conversation. It reads across months of conversations, not only the one that just ended:

• A follow-up draft after each client meeting: runs when a client meeting is detected.
• A spec from a brainstorm: runs when you ask, after the session.
• Prep for a recurring meeting: runs the morning of the meeting.
• Every open commitment, across everyone: runs Sunday evening.

The output comes back as Markdown or PDF, or as a custom section in the daily report you already read. Because Pastok is connected over MCP, your other assistant can read that report too. The Pastok setup guide covers Chat and Skills in more detail.

Let the assistant draft, and do the sending, booking, and paying yourself. Before you send something, check the parts that would be expensive to get wrong: names, dates, amounts, and who agreed to what. When a claim matters, ask ‘Where did that come from?’ A good setup points you to the email, document, or conversation behind it. For anything said out loud, the exact words are in the transcript in the Pastok web app.

Keep a short note of what you had to fix. If you're correcting the same thing every week, it belongs in the instruction.