Googlebook's Magic Pointer brings a big change:
Previously, when using AI,
you often had to:
Copy.
Open Gemini.
Paste.
Then explain:
“Here’s what I’m currently looking at.”
Magic Pointer directly brings Gemini to:
the exact place you’re pointing at.
So prompts can be much shorter.
But this also brings a new problem:
You move the cursor a bit,
and only say:
“Handle this for me.”
AI can see the context,
but it doesn’t necessarily know:
Which specific part you want to focus on?
How far you want it to proceed?
Which steps should be skipped for now?
Today, let's learn a simple method:
Before calling Magic Pointer each time,
add these three clarifications:
Where to look.
What you want.
What not to do.
The first sentence: Where to look
Google’s official design focus for Magic Pointer
is to let Gemini understand exactly:
the text,
images,
and related context you’re currently pointing at.
Google DeepMind even calls this interaction concept:
Show and Tell.
Before, you had to write:
“Please look at the date in the third paragraph in the middle of the webpage.”
Now, you can:
simply select it.
So the first step isn’t:
writing a longer prompt,
but rather:
selecting only the truly needed content.
For example, if you’re viewing a:
Marathon Training Plan
that includes:
training dates,
mileages,
diet suggestions,
advertisements,
and other article links,
but you only need:
the running dates,
then first select:
just the area containing the dates and training details.
Don’t start by treating the entire page as the task.
The second sentence: What you want
After selecting the context,
tell Gemini clearly:
what specific outcome you want.
For example, instead of just saying:
“Help me organize this,”
say:
“Only organize into a list of dates and training items.”
The key here isn’t about writing a fancy prompt,
but about giving your output a clear:
destination.
For example:
summarize into 5 points,
list 3 dates,
compare two products,
identify suspicious signals,
translate the selection into Traditional Chinese,
or organize into a to-do list.
Once AI already knows “what this is,”
you need to supplement with:
“What I want it turned into.”
The third sentence: What not to do
This is the easiest part to skip.
Google officially demonstrates:
Magic Pointer can select the Marathon Training Plan,
then let Gemini add those training sessions:
into Google Calendar.
This means it doesn’t just:
read;
it might also:
act.
If you actually just want to check dates first,
you can add a line like:
“Don’t add to Calendar yet.”
This is the third layer:
what not to do.
Putting the three sentences together makes your instructions clear
Don’t say:
“Handle this for me.”
Instead, say:
Where to look:
Only look at the training dates I’ve currently selected.
What you want:
Organize into a list of dates and training items.
What not to do:
Don’t add to Calendar yet.
It’s that simple.
Why Magic Pointer needs these three clarifications even more
Because the biggest trouble with traditional chatbots is:
insufficient context.
So you keep including more data.
Magic Pointer aims to solve this.
Google DeepMind’s design direction is to:
let the pointer understand exactly what you are pointing to,
such as:
a single word,
a paragraph,
a portion of an image,
a code block.
You can even say:
“Fix this.”
“Move that.”
without describing the whole world every time.
This is very convenient.
But when prompts get shorter,
another ambiguity starts to matter:
How far do you want it to go before stopping?
Example 1: Spotting a suspicious Email
Google’s official demo shows:
You can hover over a suspicious email,
letting Gemini analyze simultaneously:
text,
images,
and decide if it’s more like:
spam
or a regular email.
Don’t just ask:
“What should I do with this email?”
Instead, say:
Where to look:
Analyze only the content and images in this email.
What you want:
List the top 3 suspicious signals.
What not to do:
Don’t reply, don’t click links, don’t take any external actions for me.
First get:
the assessment materials.
Then decide yourself to:
delete it,
report it,
contact the sender,
or take other steps.
Example 2: Comparing two images
Magic Pointer can also:
select several images,
and let Gemini visualize them together.
For example,
you’re looking at a photo of your living room,
and next to it is a picture of a new sofa.
You could say:
Where to look:
Use only the living room and sofa images I selected.
What you want:
Generate a preview of the sofa placed in the living room.
What not to do:
Don’t alter the walls, floor, or other furniture.
Here, “what not to do”
doesn’t necessarily mean:
don’t pay,
don’t send an email,
it can simply mean:
don’t change anything I’m already happy with.
Example 3: Looking at multiple project dates
Suppose you’re viewing a:
event planning page
with dates for:
venue confirmation,
photography,
client review,
and official launch.
You could first:
select those four dates,
then say:
Where to look:
Only focus on the four milestones I selected.
What you want:
Sort by date, and list possible conflicts.
What not to do:
Don’t change the Calendar, and don’t notify anyone.
Here, Magic Pointer acts as a:
checking tool,
not an
automatic executor.
When running a new task for the first time, it’s best to stop at “organizing” first
For tasks you’re trying with Magic Pointer for the first time,
a safer approach is:
First round:
just read.
Second round:
organize/draft.
After confirming no issues,
third round:
execute.
For example:
View Training Plan.
↓
List the dates.
↓
Confirm yourself.
↓
Then ask to add them to Calendar.
Instead of:
first time seeing the data,
directly combining:
reading + judging + executing
all in one sentence.
Just because AI can skip a step doesn’t mean humans can skip a confirmation
Googlebook’s direction is clear.
Previously:
View webpage
↓
Copy
↓
Open AI
↓
Paste
↓
Explain context
↓
Get answer.
Magic Pointer removes many of these frictions in the middle.
It becomes:
Point
→ Ask
→ Result.
This is an efficiency gain.
But if the result then:
modifies Calendar,
generates official content,
changes images,
or takes other consequential actions,
a human gate should still exist.
In other words:
Reducing data transfer
does not equal
reducing necessary checkpoints.
Selecting a smaller scope is not a technical “privacy isolation”
Here we need to clarify an important point.
Today’s lesson about:
selecting just the content you really need
is mainly to:
make the task clearer.
Google DeepMind’s AI Pointer design uses:
visual context
and
semantic context
to understand what you’re pointing at.
So don’t misunderstand:
“I only highlighted this part”
as
a system-level guarantee that only those pixels are visible and no other context is ever involved.
Google does not currently make such a promise.
True privacy boundaries
still depend on:
real product features,
account settings,
data handling policies,
and current privacy documentation.
What this method handles is:
task scope,
not creating a new
security boundary for the product.
Magic Pointer only activates when you call it
Google’s current explanation about Magic Pointer is:
it only activates when you:
actively move the cursor to invoke it.
When not in use,
it stays off.
Googlebook also allows turning off Magic Pointer.
This means Google at least considers:
user invocation
as an important control point.
But after invoking it,
the task still depends on:
how clearly you express it,
and how clearly the boundaries are.
Remember just these three dimensions today
Next time you see Magic Pointer,
don’t just think,
“Finally, I don’t have to write prompts.”
The background for the prompt might be shorter,
but the task boundaries still need to be there.
Simply remember:
Where to look
What part does AI need to handle this time?
What you want
Do you want a summary, a comparison, a draft, or a list?
What not to do
Which step shouldn’t be executed yet?
For example, the simplest prompt:
Only look at the dates I selected, first organize into a list, don’t add to Calendar.
This is already much safer and clearer than:
“Handle this for me.”
The real time-saving aspect of AI-native computer work
should be:
no longer moving context around,
not:
removing human judgment and stopping points.
If you want to know which step in your workflow is best to hand over to AI first, leave a comment with “process.”
Today, let’s improve a bit with AI.
Learn one AI skill daily.
Save a bit of time every day.
Enhance your skills a little each day.
SasaDaily, growing with you.
