Notion AI: Custom Agent Automation
Key Points
- Notion just launched a new AI feature that lets users build “custom AI agents” by linking Notion databases with external tools, effectively turning the platform into an automation hub.
- The video outlines three parts: an overview of the release, live notes on what works and doesn’t (including prompting tips), and concrete demos such as an interview coach, turning meeting notes into product requirement docs/backlogs, and a prompt‑evaluation harness.
- Notion markets these agents as “AI‑powered agents across your Notion portfolio” that can perform multi‑step autonomous work for up to about 10‑20 minutes, while also adding connectors for Google Drive, Gmail, Linear, GitHub, and more.
- A key use‑case highlighted is creating custom agents that act like teammates—e.g., automatically turning sales contracts into technical requirements for engineering—showcasing the tool’s flexibility for cross‑department workflows.
- Notion’s broader value proposition is that by centralizing data and automation within its platform, users can cut costs on separate tools, positioning the AI agents as a money‑saving productivity boost.
Sections
- Notion AI: Easy Custom Agents - The speaker introduces Notion's newly released AI features, highlighting its seamless database integration and custom connectors that let users build flexible AI agents for tasks such as interview coaching, converting meeting notes into product backlogs, and evaluating prompts.
- Notion AI Automates CRM with MCP - The speaker explains how Notion uses Model Context Protocol (MCP) connectors to automatically pull Gmail responses and meeting transcript notes into its CRM, updating prospect status in the sales pipeline, and hints at upcoming releases that will let businesses add custom MCP connectors for deeper data integration.
- Tables First, Quality Checks - The speaker advocates prioritizing table-based data over raw text for clarity and efficiency, and stresses the importance of explicit AI quality checks to verify task completion criteria.
- Undo Button, Run Logs, Prompt Discipline - The speaker emphasizes that incorporating an undo function, detailed run‑log outputs, and strict, plain‑language prompting dramatically enhances agent reliability, auditability, and reduces hallucination.
- Structured Table Generation Prompt - The speaker outlines a detailed prompt that instructs an AI to create and manage structured table entries, perform quality checks (including length, content, banned words, and versioning), handle duplicates, and log changes.
- Notion as Prompt Evaluation Hub - The speaker outlines how to use Notion to create a database that logs, versions, and scores prompts for various AI tools, enabling systematic tracking, rubric‑based evaluation, and self‑improvement of prompts.
Full Transcript
# Notion AI: Custom Agent Automation **Source:** [https://www.youtube.com/watch?v=BP-N7xjz-vM](https://www.youtube.com/watch?v=BP-N7xjz-vM) **Duration:** 00:22:13 ## Summary - Notion just launched a new AI feature that lets users build “custom AI agents” by linking Notion databases with external tools, effectively turning the platform into an automation hub. - The video outlines three parts: an overview of the release, live notes on what works and doesn’t (including prompting tips), and concrete demos such as an interview coach, turning meeting notes into product requirement docs/backlogs, and a prompt‑evaluation harness. - Notion markets these agents as “AI‑powered agents across your Notion portfolio” that can perform multi‑step autonomous work for up to about 10‑20 minutes, while also adding connectors for Google Drive, Gmail, Linear, GitHub, and more. - A key use‑case highlighted is creating custom agents that act like teammates—e.g., automatically turning sales contracts into technical requirements for engineering—showcasing the tool’s flexibility for cross‑department workflows. - Notion’s broader value proposition is that by centralizing data and automation within its platform, users can cut costs on separate tools, positioning the AI agents as a money‑saving productivity boost. ## Sections - [00:00:00](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=0s) **Notion AI: Easy Custom Agents** - The speaker introduces Notion's newly released AI features, highlighting its seamless database integration and custom connectors that let users build flexible AI agents for tasks such as interview coaching, converting meeting notes into product backlogs, and evaluating prompts. - [00:03:31](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=211s) **Notion AI Automates CRM with MCP** - The speaker explains how Notion uses Model Context Protocol (MCP) connectors to automatically pull Gmail responses and meeting transcript notes into its CRM, updating prospect status in the sales pipeline, and hints at upcoming releases that will let businesses add custom MCP connectors for deeper data integration. - [00:07:18](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=438s) **Tables First, Quality Checks** - The speaker advocates prioritizing table-based data over raw text for clarity and efficiency, and stresses the importance of explicit AI quality checks to verify task completion criteria. - [00:10:49](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=649s) **Undo Button, Run Logs, Prompt Discipline** - The speaker emphasizes that incorporating an undo function, detailed run‑log outputs, and strict, plain‑language prompting dramatically enhances agent reliability, auditability, and reduces hallucination. - [00:14:00](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=840s) **Structured Table Generation Prompt** - The speaker outlines a detailed prompt that instructs an AI to create and manage structured table entries, perform quality checks (including length, content, banned words, and versioning), handle duplicates, and log changes. - [00:19:42](https://www.youtube.com/watch?v=BP-N7xjz-vM&t=1182s) **Notion as Prompt Evaluation Hub** - The speaker outlines how to use Notion to create a database that logs, versions, and scores prompts for various AI tools, enabling systematic tracking, rubric‑based evaluation, and self‑improvement of prompts. ## Full Transcript
I'm here to tell you about the very very
easiest custom AI agent automation
software out there today. And not a lot
of people are talking about it as if
that's what it is, but that's what it
really is. And we're going to talk about
it. It's Notion AI. They just released
this week. And the reason why I'm
calling it custom AI agents is because
of the way Notion has been able to marry
together databases and custom connectors
to other tools that you use throughout
your daily life. and also of course the
power of AI. So we're going to do this
in a couple of sections in this video.
Number one, I want to tell you what
notion released. Number two, I want to
give you my live actual notes using it,
including things that don't work well
and things that do work well, tips for
prompting, all that stuff. And number
three, I want to actually show you what
Notion can do with specific examples.
And so we're going to get into Notion as
an interview coach. We're going to get
into notion uh getting your meeting
notes into a product requirement
document into a backlog. We're going to
get into notion helping you with your
prompts as a prompt evaluation harness.
There's a lot of really cool stuff in
here and it underscores how flexible
this tool is, which is why I'm calling
it a custom AI agent builder, even
though spoiler alert, Notion did not
call it that. Okay, so what's in the
box? What did Notion claim that they
released? So what notion called this is
really an AI powered agentic future.
They talked about it really as AI agents
across your notion portfolio rather than
AI agents powering your whole workflow.
And I think that there's a really big
difference there. What they want you to
see is that notion's AI agents can
perform autonomous work across multiple
steps. They claim up to 20 minutes. When
I was testing it, I got five or 10
minutes pretty easily. and they are
adding tools to make that more useful.
So it's not just notion. So they're
trying to add other connectors as well
including Google Drive, including Gmail,
including your linear, including the
GitHub tool stack, including a bunch of
others. It's like they're trying to add
as many tools as they can. They also say
very shortly they're going to give you
the ability to have customizable agents
that act like teammates and take
specific work flows for specific
projects across departments. So imagine
you want to always take a contract from
sales and move that contract into
technical requirements for your
engineering team. They're trying to
build custom agents to solve for that
use case. And of course notion benefits
because you're pulling more of your data
into notion. And in fact, this is a
really interesting value proposition
because when you hit their landing page,
what they say is that notion saves you
money. You want to spend your money here
because notion saves you money on a
bunch of other things. That has been one
of their larger value props in the age
of AI. And I think it's going to
resonate because everybody knows that
you're not going to pay 100 bucks here
and 100 bucks there and 200 bucks the
other place just for AI. You want a
single home and notion is trying to be
that home by making your data at home in
notion with AI. We shall see. But I want
to show you some use cases that make it
pretty tempting. One of the things that
they have enabled AI to do that I think
we don't really easily see other places
is granular database row permissions. So
notion now has page level permissions
for databases. And so you can actually
have notion AI make granular uh database
controls and database changes per row.
So, for example, if you are trying to do
cold outreach to a contact for a B2B
business, you can have Notion look at
the response in Gmail, look at meeting
notes in a transcript, and then come
back and update a database row in Notion
in your CRM. That helps you understand
where that prospect is at in their
journey, and maybe you move them along
in the sales pipeline. That's the kind
of thing that they're envisioning and it
does work well for that. They are also
strongly advocating a universe of model
context protocol connectors. And so
remember how I've talked in the past
about MCP as sort of something anthropic
seedated into the ecosystem and
engineers have now picked up and used
across all of AI. That is true at notion
as well. Notion is using MCP servers and
bragging about it and they are implying
quick scale as a result. They want to
add more MCP connectors. I would
strongly expect a 3.1 or 3.2 2 release
to allow businesses to add custom
connectors with their own MCPs to pull
in yet more data because that's very
much what Notion wants to do. If you
centralize more data here, you'll be
stickier. You'll use our AI. You'll pay.
You'll stay. Let's get into the actual
experience that I had in part two of
this video. What did it actually look
like? Did it actually work? So, spoiler
alert, it worked, but it is more prompt
dependent than I think you or I would
want. And so, I tried multiple ways of
prompting the system. I tried the more
casual just make it where it was like a
one or twoline pass and I was like, just
make a database for this. It did not
work as well. The ability of the system
to understand what I wanted seems to be
somewhat dependent on really strongly
typed prompting or really strongly
structured prompting. I came up with
eight prompting rules as I ran through
these experiments that I want to share
with you here that I think are highly
correlated to successful notion
prompting. And I'm going to go a little
bit farther. Even if you're not in
Notion, these are going to be useful
prompting tips for a future where we are
building digital artifacts with AI. I
talked in a previous post about this
idea that work is changing. We are going
from a world of work where we handwrote
our artifacts like docs to a world where
artifacts are more interactive where we
can sort of produce something and we can
interact with it like an applet or we
can ask people to contribute or maybe
automatic actions are taken. Notion is
really at the forefront of this trend
especially with the idea that agents can
take action against a page and update
database rows as it goes and I'll sort
of show you how that works. But to do
that you have to apply these prompting
principles. Number one, be really clear
about where you want this tool to work.
You need to say work only on this page
and its subpages or something like that
because you don't want notion to be
broadly scoped and surpriseed elsewhere
because that is equivalent in many
people's notion wikis to an unwanted
code change. You don't want that. So
specify where it works. Number two, tell
it what done looks like. Ask for a
receipt at the end. it will listen and
say, "When finished, please add a line
at the bottom of the page, either okay
or if you're blocked, add blocked and a
reason why." This will let you know
right away in the page text itself what
it did and why. And so getting receipts
helps you to be more auditable and to
track what actually happened. I've
actually developed a prompt that shows
audit logs of previous runs right on the
page itself, which I think is important
if you're starting to make serious
changes. And I'll show you sort of how
that works. Number three, you want to be
thinking in terms of tables and
databases rather than text. If you're
creating things in databases in Notion,
you're leaning into Notion's strengths.
You're leaning into the strengths of a
lot of compute. Frankly, a lot of other
systems depend on databases, too. Tables
are much easier to sort, to operate
against, to review, to fix later, to
adjust. Whereas raw text can be
difficult to format and engage with. I
did try both with notion. I really felt
like a tables first approach was much
more useful for the kind of tasks I was
doing and I think that this is something
that we're going to see change the way
artifacts are formatted. I'm used to a
world out of Amazon with product
requirement documents that were
narratives and yes you have some tables
but you also have a lot of narrative
about the customer experience. We are
moving to a world that is videoheavy and
that also has tables. And it's a really
interesting change for someone who came
up sort of before all of that happened
in the traditional six-pager era. But
here we are. Make tables first, text
second. Principle number four, you need
to use quality checks. One of the things
that you will really thank yourself for
is if you are explicit with the agent
about the conditions under which it can
mark a task accomplished or done. So you
can have it check the length of a
particular piece of text. Is it 180
tokens? Right? Is it 180 letters? You
can have it check if it includes every
piece of data that's relevant for the
task. So if you're writing a cover
letter and you tell notion to do that,
which by the way it can have it include
the company and the role. That seems
like a reasonable requirement. If you
are writing a justification for why you
should work somewhere, make sure it
includes at least one number from your
resume. And you can give it the resume
and it will do that. And the specificity
that you can use here around quality
checks is something that people forget
about, but it's a critical part of
learning to work with agents. They need
that degree of clarity in order to know
they did a good job for you. Otherwise,
they'll just guess and they may
hallucinate or they may not do anything
at all or they may default to token
efficiency and do less than expected. So
use quality checks and also be really
clear what the model should do if it
doesn't have that item. And so you can
write if info is missing please insert
TK confirm which is a traditional
editorial slogan. Or you can pick
whatever you want. You can say insert in
brackets please check this and ask you
for it and it will do that. This helps
the model not just depend on vibes but
actually get to a past fail mindset.
Principle number five don't create
duplicates. If something similar already
exists, you want the model to update it
instead of creating new copy and
dirtying up your context window because
you're starting to think of notion and
really we're starting to think of
agentic tools as context windows
themselves. Like even if you can't
absorb all of it in the context window,
notion itself is becoming a place where
you always have that in mind because of
the way AI operates on it. And if it's a
context window, it needs to be clean,
which means you don't want duplicates.
And so you can actually specify when you
touch a page, if you updated it, please
include a little table where it has
version and last run to describe your
edit and describe the last time you
touched it. This enables you to see what
happened and see how pages changed. It's
one of the things that's going to be
increasingly important as humans and
agents work together in wikileike
environments. Number six, I want you to
create a run log. I want you to think
about each change you're making as if
it's something that needs to be undone.
One of the hallmarks of good agent
architectures is the ability to hit
undo. And I appreciate that notion has
put a literal undo button in the chat
interface. I haven't seen that a lot. I
think people are going to appreciate
that. But if you print a tiny run log,
it's going to help you to go farther.
And so you can stick that into the
prompt. It's one of those things that
sort of extends the idea of a page
update note into a full run log that
actually has links to what happened,
warnings for when things go wrong, etc.
The more you invest on the validation
and audit side, the more you can keep
your context window happy. And by the
way, if you think this is overkill, this
is just not so hard if you have the
right prompt. like I did not actually
have to suffer that hard creating the
pages I made because I was able to work
with chat GPT5 in thinking mode to
create the prompts. And so I'm going to
go through the sort of some of the sort
of conversations, the prompts, what I
got. You'll get the idea. You'll get how
I was able to use these eight principles
without too much blood, sweat, and tears
on my part. Principle number seven,
write in really plain strict language.
So, say create six questions instead of
create a few questions. Use one metric
instead of use a metric. Please,
wherever you can, you want to avoid
open-ended phrases that will encourage
the model to hallucinate unless you are
happy with hallucination. So, as an
example, be inspiring is not a helpful
frame for a cover letter. Asking it for
a specific metric, asking it to include
the name and the company, asking it to
include a specific reason from your
resume. This ensures the agent stays
consistent. So write as plainly and
strictly as you can. And GPT5 is
actually very helpful in this. You're
working with its default language
preferences. Principle number eight,
please don't let it make things up. I
know I said unless you want
hallucination, but really wherever you
can, you want to be underlining to the
model. If you cannot find a claim in the
input data that I give you, please use
check this and do not mark it as done
and come back to me because you don't
want to get into a situation where
you're making up dirty data and then the
model is basing future actions on that
dirty data. So with that in mind, let's
put it all together and let's first look
at a sample prompt in GPT5 that helps us
understand how notion works and then
look at some notion pages that I was
able to create with that kind of a okay
here we are. We have obviously the role
you are a notion agent you want to
specify and limit the page and the
subpages here and you want to make it
clear what done looks like. This is
where I include showing receipts. what I
where I include showing what blocked is
and why. You also want to make sure that
you define the scope. I do not want it
touching things that were sort of
created or edited a long time ago.
Again, I'm trying to keep this context
window as clean as possible. Um, please
do not overwrite unless you are updating
a newer version. And so, it's very
precise about when overwrites happen and
why. Uh, this is a table and I give it
the choice to create or not. So, I could
have this table already here or not. If
it's not here, I tell it to create it.
And I'm literally giving it the columns
and I'm showing it what I want and what
the format is of each column. It's very
specific. Name, which is a title. You
have notes, which are text, a version,
which is a number, etc. Then I get to
the tasks the model should do. Find up
to five items that need work. You can
see we're starting to build a to-do list
here. For each item, draft the content
in the table fields. Run quality checks.
See below. And you tell it to look down.
If all the checks pass, set the status
to ready. If the checks fail, set the
status to needs fixed. Quality checks
then include length is within limits,
company enroll if relevant, at least one
number, avoids banned words, if info is
missing, insert. By the way, avoids
banned words will help you. If you are
writing for a quote unquote AI detecting
uh tool, like there are some tools now
that companies use that others use that
claim to detect AI wording. you can
pretty easily game them if you come up
with a list of words that AI tends to
use like delve and make sure that it
doesn't right and so there's ways you
can start to even game the writing style
here. Then it gets into duplicates and
versions how you handle update what the
requirement is. It's last 10 minutes in
this case. Uh it gives me a version
number and then finally please add a row
in the run log with the times and the
items changed and the warnings. And so
as complicated as those eight principles
sounded, I got all of that fixed into
about a 20line prompt and it's
relatively easy to run. Let's see what
it looks like in practice on a few
actual pages. All right, here we are. We
are looking at the meeting notes to PRD
backlog. I constructed this in just a
couple of minutes. As long as you have
the data, you can do that, too. Uh you
might be wondering, how did I do this?
This looks really complicated. There's
multiple tables here. They scroll along.
You can see that these tables have
statuses that have now changed. You have
PRDs. If I click the PRD, I'm actually
going to see a real page here. So, let
me just click that. Um, and I can click
it and kind of look in and see where the
PRD is. Oh, here we go. Um, so it gives
me an acceptance test, a goal, a
problem. It's actually writing the PRD
as a table, which is really cool because
then it can do operations against
individual components of that PRD. This
is a great example actually of an
artifact that is created to be agent
readable first and human readable second
because it's very easy for me to look
and say what is the what is the TLDDR of
notion agents reliability and I can get
a nice summary from notion. Um and so if
I go here and if I like copy and paste
this
um let me just pull up the actual notion
you can see where I actually did this.
uh please give me a 20word summary of
this PRD and it will just come back and
it will work on on doing that as we chat
here it's looking at it thinking about
it and there you go that's what's in the
box uh what is the highest risk element
of this build right I can actually then
start to inquire into how it works and
so that's one of the powerful things
here right like you can start to
actually ask it to exercise judgment ask
it to think through is talking about
schema drift here you may or may not
agree But the point is you can have that
conversation with it very easily. Now if
we go down here, it can actually go and
automatically fill out to-dos associated
with these PRDs. And so these are all
associated with particular PRDS. And
these are basically to-dos for the data
team, for the email team, the backend
team, all automatically created. All I
did to add to this was to input some
data from meetings. And you can actually
even automate that because notion now
has meeting notes that you can take by
audio. And so I'll share the the
original prompt that I got for this um
and you can sort of see how you can make
it your own. But it illustrates to me
that it's increasingly possible to
move from a world where you consider
these artifacts as static to one where
they are truly dynamic and you can
actually like evaluate how the overall
projects break out in a matter of a
couple of minutes rather than a couple
of days or even longer. I remember when
I was doing PRD work as a product
person, what I'm showing you here would
take days. It took about 2 minutes and I
think that's really compelling. Let me
show you another cool example that I
found. So, this is the notion interview
coach. It may not look like a lot, but
it gives you a rubric and it gives you
everything you need to actually run your
own notion interview scorecard. And so
this is actually it's simulated data,
but what you see here is an entire
database where it can take a notes
transcript. Let's say you interview
yourself and you're practicing your
answers. It can take a notes transcript
with questions, feedback, interviews,
and it can put it into a database like
this and it can actually run against a
rubric for clarity, impact, specific
specificity, structure, whatever you
want. score it and deliver an overall
scorecard to you of how you did. Um, and
I think that's really cool. It was
relatively easy to spin up. There's a
lot more we could do, but it shows you
that you can actually build an entire
system with multiple databases off of a
single prompt and start to actually work
to populate it with real data and get it
going from there. Let me show you one
more and I think you'll find that really
sort of an overall picture of what
notion can do. My goal here is not to
give you the complete picture of notion
because I don't think I can do that. I
want to give you a sense of how I think
notion undersold this. This is actually
an agentic artifacts factory. There's a
lot more to this and I think that with
proper prompting you can go a long way.
So let's do one more. So this is a
prompt and eval harness. This is going
to be more technical and you can
actually look at particular experiments
that were run. You can look at a uh date
for those experiments. You can look at
inputs. You can look at versioning
essentially. And then you can look at
results that were scored pass or fail
based on a rubric, right? Um and you can
get into eval rules. What must the
prompt include? Did the prompt work? Did
it not work? A run log on updates. Um,
at the end of the day, I think what you
should be taking away from this is that
you can do things as nerdy as
self-improving your notion prompts by
using Notion AI. You can do things as
detailed as getting into particular
prompt structures for different tools,
say your perplexity prompts, your OpenAI
prompts, your Claude prompts, and start
to track them in a thoughtful way in a
database. I've been, you know, told by a
lot of people that they are looking for
great prompt tooling. And there's lots
of answers to that question depending on
your workflow. But one of the answers is
notion. One of the answers is actually
building out a prompt database in
notion. And I'll share this prompt in
the in the post and starting to track
and score how your prompts do. Now, if
you're one of those people who says, "I
don't care about prompts." That's fine.
But you're probably going to be getting
better results if you take your prompts
this seriously and actually start to
score them. And of course you can adjust
them to score how you want. This is just
a sample score. Uh and you can see how
the sample score works. Um yeah, so this
is one of those things that I think is
getting slept on and I'm sharing about
notion because I think that we need to
get past an assumption that work is a
series of individual things that we
create with the help of AI and we need
to move to a to an idea of an
agentpowered work factory where the
agents are processing through these
artifacts often autonomously and it's
our job to prepare the environment and
to shape the direction for these agents.
And that sounds super fancy and it
sounds like a big company thing, but
Notion is making that possible for
anybody. Notion is making that possible
at a price of like 20 bucks a month.
Like it it's really very affordable to
have this kind of a thing. And I think
that's really really cool. And so I hope
you've enjoyed this breakdown of notion.
I hope you see why I think it's really
interesting. We are headed to this
future where agents are powering
artifacts. I hope these prompts that I'm
going to share are helpful to you as
well. Cheers.