Every AI assistant is generic until you tell it otherwise. It does not know your job, your audience, the words your company bans, or the format your boss wants. So it guesses, and you spend the next ten minutes correcting it.
A personalized AI assistant fixes that once, instead of in every chat. You do not need a special tool. You need four layers of information, each in the right place.
The four layers
| Layer | What it holds | Changes |
|---|---|---|
| Who you are | Role, organisation, audience, expertise | Rarely |
| What you do | The recurring tasks, each with its own brief | Quarterly |
| How you want it | Format, tone, and the never list | When something annoys you |
| What it can check | Style guides, templates, past examples | When the source changes |
Where each layer goes
| Who and how | What you do and files | |
|---|---|---|
| ChatGPT | Custom instructions | A Project per workstream |
| Claude | Profile preferences | A Project per workstream |
| Gemini | Saved info | A Gem per task |
| Microsoft 365 Copilot | Personalisation settings | An agent per task |
The principle is the same everywhere: things true in every chat go in the account-level setting; things true for one kind of work go in a project, Gem or agent.
A starter setup you can copy
Account level: who you are and how you want it
I am a [role] at [organisation]. My work is read by [audience],
who care most about [what they care about].
Answer first, then reasoning. Plain language. If something depends
on a detail I have not given, ask one question before answering.
Never: preamble, praise for my question, invented numbers, or a
closing summary.One project for your most common task
This project is for [task, e.g. weekly client reports].
Every output must: [the three things a good one always has].
Format: [structure, length].
Use the attached [template / examples] as the standard.
If information is missing, list it at the top instead of filling
the gap.Then attach two or three examples of the task done well. Examples teach tone faster than any description.
What goes wrong
- One giant instruction block. Everything in the account setting means the rules for reports leak into your emails.
- Twelve projects on day one. Build one, use it for two weeks, then add the next.
- Relying on memory. Memory stores what you happened to mention, not what you decided matters. Write the important things down on purpose.
- No examples. Describing your style in adjectives rarely works. Showing it does.
Questions people ask
Which AI is best as a personal assistant?
The one your employer allows and you will open every day. Setup matters more than the model. A well-configured assistant in any of the four major tools beats an empty one in the "best" tool.
Is it safe to give an AI assistant my work files?
Follow your company's AI policy. As a rule, use work accounts for work content and keep customer personal data out of files you upload.
How long does setup take?
About an hour for the account level and your first project. The payback shows up within the first week.
Where to take this next
Start with the account-level block above and one project. If you want all four layers written around your actual role, The Pack builds them from your LinkedIn profile: instructions plus eight task files, with install steps for ChatGPT, Claude, Gemini and Copilot.
Laisser un commentaire :