Planning and organising
Appointments, lists, trips and the household: where AI genuinely helps with organising, where it gets in the way, and what not to hand it.
Where it helps
| Task | Any good? |
|---|---|
| Drafting a packing list for a trip | very good |
| Building a weekly menu with a shopping list | very good |
| Breaking a project into steps | good |
| Planning a party with a schedule | good, check the timings |
| Finding times and computing travel | poor |
| Stating opening hours or prices | poor, out of date |
| Computing binding deadlines | do not use |
An example
"I am travelling with two children, four and seven, for a week in October to South Tyrol. We will hike, it may rain. Produce a packing list, grouped by person, and mark what we probably already own."
The result is a list you cut and extend. That is exactly the value: the blank page is gone, and what is missing becomes obvious as you read.
What makes the difference
- 01
State the constraints
How many people, how much time, what budget, what is ruled out. Without constraints you get an average plan.
- 02
Specify the format
Table, list, grouped by day. That saves the rework.
- 03
Ask what is missing
"What is missing from this list that people typically forget?" That question often yields more than the first draft.
- 04
Decide yourself
The plan is a proposal. The order that suits you is not something the model knows.
When an assistant reaches into your data
Calendar, mail and tasks can be connected. That changes the question fundamentally: no longer what you type in, but what a provider sees permanently.
- Which permissions are requested, and are they necessary for the purpose? Read access to everything rarely is.
- Is content stored, and for how long?
- Is content used for training?
- Can access be withdrawn again, and what happens to what was stored?
Why date arithmetic is hard
A language model handles numbers as pieces of text, not as values, see Tokenisation. Date arithmetic requires calendar rules, time zones, daylight saving and holidays per region. None of that is present as a rule, only as a pattern in text. The result is plausible and frequently a day or an hour out.
- Hand date arithmetic to a tool that computes, not to the model.
- Always have the time zone written out with a proposed time; then the error becomes visible.
- Never take a deadline with legal effect from a model output.
Assistants that act
As soon as a system does not merely propose but creates appointments, sends mail or orders things, the consequence of an error changes. A wrong proposal costs nothing; a wrong action does.
- Confirmation before any action that reaches outside or costs money.
- A hard limit on the number of steps per instruction.
- Separate permissions per tool instead of one credential for everything.
- No access to content that could contain instructions without a check. See Agents in depth.
The last point is the underrated one: a mail from outside can contain an instruction aimed at the assistant. A system that reads foreign content and is also allowed to act is exposed.
The planning template
FREE ACCOUNT
Templates for a trip, a week and a project
Three templates with the constraints that make the difference, and the question to ask after the first draft.
Templates2 items