Build a Smart Personal Assistant

What is a personal AI assistant

A personal AI assistant isn't just another chat — it's a chat model (like ChatGPT or Claude) that you set up in advance to know you: who you are, your goals, the tone to answer in, and its role. Instead of starting every conversation from scratch — the assistant already knows, and is available 24/7 for any task.

Why use a personal assistant

• Real time savings. No re-explaining context every time — jump straight to work.
• Consistency. Answers and outputs in your voice and style, always.
• Always available. 2am, weekends — it's there, no queue and no fatigue.
• Less mental load. It breaks down decisions and tasks so your head is free to think.

What a personal assistant can do

• Drafting & writing: emails, messages, posts, summaries — ready to copy, in your tone.
• Thinking & decisions: brainstorming, pros vs cons, analyzing options.
• Organizing & planning: breaking a project into tasks, prioritizing, planning day and week.
• Learning: explaining topics at levels, summarizing, practice questions.
• Checking & polishing: proofreading, improving phrasing, spotting logic errors.

How to build it — step by step

1. Pick a platform — ChatGPT, Claude or Gemini (see below what fits).
2. Set a "system prompt" — the text that defines who you are, how it answers and its role. This is the heart.
3. Fill in your details — name, field, goals, tools, and preferred tone.
4. Save it — as a Custom GPT (ChatGPT), Project (Claude), or Gem (Gemini) so it stays.
5. Refine as you go — when an answer misses, update the prompt. Within a week it gets sharp.

The full prompt (copy & adapt)

This is the system prompt that turns a regular chat into a personal assistant. Replace the brackets with your details:

You are my smart personal assistant. Your job is to save me time, think with me, and help me get things done efficiently.

# Who I am
- Name: [your name]
- Field/role: [e.g. small-business owner / student / marketing manager]
- Main goals: [2-3 goals]
- Tools I use: [e.g. Gmail, Google Calendar, Notion]

# How to answer me
- Be clear, direct and friendly. No fluff.
- For complex tasks, break them into actionable steps.
- When you lack info to answer well, ask me one focused question first.
- Give a short answer first, then detail if needed.
- If I ask for text (email, post, message), write it ready to copy, in the tone I asked for.

# What you do for me
- Drafting: emails, messages, posts, summaries — in my voice.
- Thinking: brainstorming, pros/cons, decisions.
- Organizing: breaking projects into tasks, prioritizing, planning the day/week.
- Learning: explaining topics at different levels, summarizing, practice questions.
- Checking: proofreading, improving phrasing, spotting logic errors.

# Principles
- Accuracy over speed: don't invent facts. If unsure, say so.
- Keep my context throughout the conversation.
- Take initiative: if there's a better way, suggest it.

At the end of any complex answer, ask: "Want me to continue / refine / take the next step?"

Which platform to choose

• ChatGPT (Custom GPT): easiest to build a saved assistant with a name, icon, files and tools. Great for general use.
• Claude (Projects): strong at long-form writing, document analysis and context memory. Excellent if you work with lots of text.
• Gemini (Gems): connected to Google services (Gmail, Docs, Calendar) — handy if you live in Google's world.

Why most of these get abandoned in a fortnight

Setting up a personal assistant is enjoyable — you write the prompt, give it a name, test it on something, and it works. Then two weeks later you are back to opening a blank chat.

The reason is worth stating plainly, because it determines what the rest of this page should be about: the setup is the small part. What makes an assistant genuinely useful is a habit of going to it first, and a habit forms only if the thing is noticeably better than the blank chat on the work you actually do.

Three things make that difference, and only one of them is the system prompt:

• It knows the specifics of your situation, not just your job title. This is context you supply, and it is most of the value.
• It is where your work already is. An assistant in a tab you never open loses to the one in the app you have open anyway.
• You improve it when it disappoints you, rather than shrugging. Ten small corrections over a month produce something that feels tailored; nobody gets there in one sitting.

So treat the prompt below as a starting point rather than a finished product, and expect the good version to exist in about a month.

The context file does more than the persona

Most people write a persona — "you are a professional assistant with expertise in marketing" — and stop. That paragraph does very little. What changes the output is the boring, specific material about your actual circumstances.

Keep it in a document you own, and attach or paste it. Roughly:

• What you do, concretely. Not "marketing" but "I run marketing for a six-person B2B software company selling to accountancy practices."
• Who you write to. Your audiences, and how formal each expects you to be.
• What you are working on right now — the two or three live projects, with enough detail that a reference to "the migration" lands.
• Your constraints. Budget, team size, tools you must use, things you cannot change.
• How you want it to write. Sentence length, whether you use contractions, vocabulary you hate. Specific bans work far better than adjectives.
• Standing decisions. "We do not do cold outreach." "Always price in euros." Saves correcting the same thing weekly.

Keeping this as a file rather than only inside one platform's memory has a practical advantage: it is portable. When you switch tools, or set up a second assistant for a different job, you paste the same file in and you are immediately back to a useful starting point.

Update it when your situation changes rather than when you remember. A context file describing last quarter's priorities quietly produces last quarter's advice.

Which parts of a system prompt actually work

Prompt templates circulate with a lot of decoration in them. From the parts that recur, these earn their place:

• The role, stated narrowly. "Help me draft and think, not to make decisions for me" shapes behaviour. "You are a world-class expert" does almost nothing — it is the model's default register anyway.
• An instruction to ask before assuming. The single highest-value line on this page, covered below.
• Output format defaults. "Default to short. Bullet points only when there is a genuine list. No preamble, no summary at the end." Removes the most common irritation permanently.
• Explicit permission to disagree. "If you think I am wrong, say so first and explain why." Without it you get agreement, which is useless when you wanted a check.
• A rule about uncertainty. "If you are not sure, say so rather than producing a confident answer. Never invent statistics, names or sources."
• Vocabulary bans. A short list of words you never want to see. Cheap and effective.

What to leave out: long lists of capabilities it already has, flattery about how excellent it is, and elaborate personas. They consume the instruction space and change little.

Making it ask instead of guess

This deserves its own section because it changes the experience more than anything else you can add.

By default these models answer whatever you asked with whatever they have. Give an under-specified request and you get a plausible, generic response built on assumptions you did not make. You then rewrite the request, and repeat.

The instruction that fixes it: "Before answering anything substantial, ask me up to three questions if the answer would change materially depending on the answers." The bounded number matters — an unbounded version produces an interrogation.

It feels slower and it is faster. Three questions and one good draft beats four rounds of correcting a draft built on the wrong assumption — and the questions themselves are often useful, because they surface a decision you had not noticed you were making.

A variant worth adding for bigger tasks: "Summarise what you understand the task to be, then wait for me to confirm before doing it." Thirty seconds that prevents a long answer to the wrong question.

Ask for the disagreement

These models are trained to be helpful and agreeable, which produces a specific failure: ask "is this a good idea?" and you will usually be told it is, with supporting reasons.

That is worse than useless when you were looking for a check, because agreement feels like validation. Three ways to get something real:

• Ask for the opposite explicitly. "Give me the three strongest arguments against this" produces genuinely different material from "what do you think".
• Hide your position. "Someone proposed X. What are the problems?" gets a more honest answer than "I am planning to do X, is it sensible?"
• Ask what would have to be true. "What would have to be true for this to fail?" is the most useful single question you can put to an assistant about a plan.

Put a standing instruction in the prompt as well, and then notice whether it is being followed — if every answer still agrees with you, the instruction is not working and the assistant is not doing the job you set it.

Where it genuinely earns its place

Being honest about this makes the habit stick, because a tool that helps with some things and is forced onto others gets abandoned.

Consistently worth it: the blank page, on anything routine; turning messy notes into something structured; explaining a topic at the level you need; a first pass of criticism on your own draft; rewriting something for a different audience; breaking a vague project into actual next steps; translating; the email you have been avoiding.

Sometimes worth it: research where you can check the sources; planning where you supply the constraints; analysis of material you paste in.

Rarely worth it: anything where finding the answer takes less time than describing the question; decisions that hinge on things only you know; short high-stakes text where the register is the message; anything you cannot verify and cannot afford to get wrong.

The pattern is the same one that applies everywhere: it is strong when the material is in front of it and the job is transformation, weaker when it has to supply facts from memory.

What not to put in it

A personal assistant accumulates a picture of your work, which is the point and also the thing to be deliberate about.

• Never: passwords, card details, API keys, identity numbers. There is no use case that justifies it.
• Check first: client material, unreleased work, anything covered by an agreement. Whether it is acceptable depends on the platform's terms and your client's — and it is a question to ask before, not after.
• Know the training setting. Most consumer services offer a switch controlling whether your conversations are used to improve models; business tiers usually have it off by default. Find it once.
• Remember it is stored. Even with training off, conversations sit on a server for some period. Treat the chat like an email to a supplier, not like a private notebook.
• Be careful with other people's data. A colleague's performance, a customer's complaint — those are about someone who did not choose this.

One assistant or several

The instinct after the first one works is to build ten. Usually a mistake, because each needs maintaining and most of them get used twice.

Split when the context is genuinely different — a work assistant and a personal one, or one per client whose details must not mix. Do not split by task type; one assistant with a good context file handles drafting, planning and analysis perfectly well, and three specialists mean three prompts to keep current.

If you do run several, keep a single shared context file and add the specific part per assistant. Otherwise the same fact ends up stated three ways and updated in one.

Keeping it sharp

Assistants degrade in a specific way: the prompt accumulates. Every irritation adds a rule, and after a few months there are twenty instructions, several contradicting each other, and the behaviour is worse than it was at ten.

A short maintenance habit prevents it:

• When it disappoints you, fix the prompt then — not later. Thirty seconds while you remember what went wrong.
• Prune quarterly. Read the whole prompt and delete rules that address something that no longer happens. Shorter prompts are followed more reliably.
• Watch for contradictions. "Be concise" and "always explain your reasoning" fight each other, and the model resolves it unpredictably.
• Re-read the context file when a project ends. Stale context is the commonest cause of advice that feels slightly off.

Giving it your documents and connecting your tools

Every major platform now lets an assistant read files you upload, and increasingly connect to services you use — mail, calendar, storage. This is where a personal assistant stops being a better blank page, and it brings two things worth understanding.

Uploaded files beat described situations. Attaching the actual brief, the actual spreadsheet, the actual thread produces answers grounded in your material rather than in what the model assumes about situations like yours. If you find yourself explaining a document, upload it instead.

Connections change the risk profile. An assistant that can read your inbox can be influenced by what is in your inbox — a message containing instructions is content the assistant may act on. That is a real and documented class of problem, and the sensible precautions are ordinary ones: grant the narrowest access that does the job, prefer read-only where you can, and keep a human approval in front of anything that sends, pays or deletes.

A useful ordering. Start with uploads, which are explicit and bounded — you decide what it sees, each time. Add read-only connections when re-uploading the same things becomes the friction. Add write access last, if at all, and only where you would be comfortable with the occasional mistake.

What it cannot do

• Know anything you have not told it, including what happened at yesterday's meeting or what your colleague meant.
• Take responsibility. The output is yours the moment you send it.
• Remember reliably across conversations, unless the platform has a memory feature and you have checked what it actually stored.
• Hold a position. Push back and it will often fold, whether or not you were right.
• Replace thinking about something difficult. It is good at helping you think and poor at doing it for you, and the difference matters most on exactly the decisions you would most like to hand over.

Tips & common mistakes

• Give rich context. The more the assistant knows about you and your goals, the more useful it is.
• Define the tone explicitly. "Friendly", "matter-of-fact", "short" — otherwise you get a generic default.
• Ask it to ask. A good prompt tells the assistant to ask when info is missing — instead of guessing.
• Verify important facts. An assistant helps you think, but sometimes errs. Check critical data.
• Privacy. Don't share sensitive info (passwords, card details) in the chat.