[00:00]
Most people learn how to prompt Claude
[00:01]
from random tips on the internet. But
[00:03]
Anthropic, the company that actually
[00:05]
built Claude, released their official
[00:06]
prompting guide. Almost no one's read
[00:08]
it. I did. And here are the seven most
[00:11]
important rules straight from the people
[00:13]
who created the model itself. Let's get
[00:15]
into it. So, as I mentioned, these tips
[00:17]
came directly from Anthropic, the
[00:19]
company that made Claude. And this here
[00:20]
is the blog post that I'm referring to.
[00:22]
So, it walks through a series of tips,
[00:23]
not just seven, many more. But I wanted
[00:25]
to call out the seven that I feel like
[00:26]
are applicable to the broadest audience
[00:28]
where people can get direct impact of
[00:30]
using these tips. And we'll start with
[00:31]
tip one, which is to be clear. Now, this
[00:35]
is more critical than ever for all the
[00:37]
models that are relevant today, such as
[00:38]
Gemini 3, GBT 5.2, Claude Opus 4.5. All
[00:42]
of these models are very good at
[00:44]
following instructions. So, when you're
[00:46]
clear with the model, you're going to
[00:47]
get a higher quality response, and the
[00:49]
AI is going to work in the direction
[00:50]
that you care about. Vague prompts tend
[00:52]
to lead to common patterns that are kind
[00:55]
of considered AI slop today. So it's
[00:57]
generic outputs that aren't very useful.
[00:59]
Now why does this happen? It happens
[01:02]
because most models have a gravitational
[01:05]
pull towards what's most common. So here
[01:08]
we have gravity pulling in all the
[01:11]
commonalities of the internet. And
[01:12]
here's the default state. An example
[01:14]
here is UIs. So often times when an AI
[01:17]
creates a UI for you, it's going to have
[01:19]
purple gradients and it's going to have
[01:20]
rounded corners. Reason being is the AI
[01:23]
was trained on millions of examples that
[01:24]
had that type of aesthetic. So the AI
[01:27]
thinks that if it provides you that
[01:28]
aesthetic, you're going to be satisfied
[01:30]
as a user if you don't provide it very
[01:32]
specific instructions. And that's the
[01:34]
key here. We have to be specific to
[01:37]
escape the gravitational pool that is
[01:39]
all those examples the AI was trained on
[01:41]
historically. And that's what this tip
[01:42]
highlights. So I'm going to walk you
[01:43]
through a few examples of what good
[01:44]
looks like. So on the left hand side we
[01:46]
have a bad prompt. This is weak. So
[01:48]
we're basically stating create an
[01:49]
analytical dashboard. We we gave no
[01:51]
specificity here. A better prompt is
[01:54]
create an analytical dashboard. Include
[01:56]
as many relevant features and
[01:57]
interactions as possible. Go beyond the
[01:59]
basics to create a fully featured
[02:00]
implementation. And there's a few key
[02:02]
things I want to call out here. So first
[02:03]
off, we're stating I want you to include
[02:05]
as many. So be as inclusive as possible
[02:07]
when it comes to the features and
[02:09]
interactions relevant to the dashboard
[02:10]
we're trying to create. Another thing
[02:12]
we're making explicit is the AI needs to
[02:14]
go beyond the basics. So, it's going
[02:16]
beyond what's normal and creating
[02:18]
something that's fully featured that's
[02:20]
relevant to the initial dashboard we're
[02:22]
trying to create. And this is what a
[02:23]
good prompt looks like that's more
[02:24]
clear. And here's another example for
[02:26]
presentations. So, here we have a week
[02:28]
example that says create presentation.
[02:30]
An improved version of this is to create
[02:31]
a professional presentation in relation
[02:33]
to our quarterly results. I want you to
[02:35]
include thoughtful design elements,
[02:38]
visual hierarchy, and engaging
[02:39]
animations where appropriate. And as you
[02:41]
can see, we're being very clear on what
[02:42]
we want the AI to include in the visuals
[02:44]
here, and that's key. So, this is our
[02:46]
first tip, which is be clear, quick
[02:48]
pause in your regular programming. This
[02:49]
video is brought to you by me, as
[02:51]
always. So, two quick things. First off,
[02:54]
below is a 30-day AI insight series,
[02:56]
completely free. You'll get 30 insights
[02:58]
in your inbox how you can apply AI to
[02:59]
your business and your work. The second
[03:01]
thing is, if you'd like to work with me,
[03:02]
there's a variety of ways you can do
[03:04]
that, either through a private AI
[03:05]
community or directly one-on-one. You
[03:07]
can check that out below as well. Let's
[03:08]
get back into the video. Our second tip
[03:10]
is often something that a lot of people
[03:12]
skip, which is explain why. And what do
[03:14]
we mean by this? Well, when you give the
[03:16]
AI some sort of context, so you ask it
[03:18]
to do something. If you also add the
[03:20]
intent as to why you're asking the AI to
[03:22]
do this and why it's important for the
[03:24]
AI to take that task on, it can actually
[03:26]
infer a lot from what you've stated,
[03:28]
even if you've not made it explicit. In
[03:30]
other words, Claude can figure out
[03:32]
things that you didn't even say as long
[03:33]
as you're explicit about why you're
[03:35]
asking the AI to do this in the first
[03:36]
place. And that often usually deres a
[03:39]
higher quality output for you. Now let's
[03:41]
look at some examples. So this first
[03:43]
example here is a weak example. Not a
[03:45]
great prompt. We're saying write this in
[03:46]
a formal tone. It can be improved. A
[03:48]
better way of saying this is I want you
[03:50]
to write this in a formal tone because
[03:52]
remember we're emphasizing the why here.
[03:54]
It's going to our board of directors and
[03:56]
we need to look credible and
[03:57]
professional. By providing just this
[04:00]
context, the AI is going to be able to
[04:01]
create something that's more
[04:02]
aesthetically pleasing and aligned with
[04:04]
the quality you're seeking by knowing
[04:06]
this specific intent. Now, here's
[04:07]
another example. So, here we have the
[04:09]
weak prompt of saying, "Keep it short."
[04:10]
A better version of this is stating,
[04:12]
"Keep it short because, again, we're
[04:13]
highlighting the Y. I'm sending this to
[04:15]
my team via text message, and longer
[04:17]
messages don't get read." Again, we're
[04:19]
highlighting the Y, which is going to
[04:20]
get better quality outputs from the AI.
[04:22]
That's our second tip. And now, we'll
[04:24]
move on to tip three, which is giving
[04:25]
good examples. So the thing with
[04:27]
examples, especially in relation to the
[04:29]
state-of-the-art models today, such as
[04:30]
Opus 4.5, Gemini 3, GPT5.2,
[04:34]
all of these models, they follow the
[04:36]
examples to the tea. So if you include
[04:37]
an example in your system prompt, the AI
[04:39]
is going to follow that very
[04:40]
specifically. So it's important that you
[04:41]
don't misdirect the AI by accidentally
[04:43]
including something in your example that
[04:45]
you don't want in its output. And a meta
[04:47]
lesson in relation to this tip is not
[04:49]
only is the example that you explicitly
[04:50]
call in the prompt dictating what the AI
[04:52]
gives you, but also the way that you
[04:54]
write your prompt determines how the AI
[04:56]
gives you an output. What does that
[04:57]
mean? In the sense that if I write my
[04:59]
prompt in a very playful, fun, and
[05:01]
simple way, the AI is likely going to
[05:03]
respond and give me an output in that
[05:04]
way. If I might write my prompt in a
[05:06]
very formal and structured way, the AI
[05:08]
is going to do the same. So even the way
[05:10]
that I write my prompt is going to
[05:12]
dictate how the AI outputs its response
[05:14]
to me. So it's important to note not
[05:16]
just the example but also your prompt
[05:17]
matter. Now on to tip four which is
[05:19]
asking for the format. Not saying
[05:21]
something that you don't want. You want
[05:22]
to say what you do want because often
[05:24]
when you look at system prompts people
[05:26]
have really huge negating terms of
[05:28]
saying never do this or don't do that or
[05:30]
whatever else. Well with AI today since
[05:32]
it follows instructions so well all you
[05:34]
have to do is say what you want instead
[05:36]
of what you don't want and the AI is
[05:38]
likely going to do that more
[05:38]
effectively. What are some examples of
[05:40]
this? Well here's an example. There's
[05:42]
first one is a weak example. So we say
[05:44]
do not use markdown in this response.
[05:45]
The reason this is weak is because we're
[05:47]
using a negating term. So we're saying
[05:48]
do not. A better version of this is your
[05:52]
response should be composed of smoothly
[05:54]
flowing pros and paragraphs. This is
[05:56]
going to give us what we want instead of
[05:58]
avoiding what we don't want. Another
[05:59]
example is here we have a weak version
[06:01]
saying make it look nice, not very uh
[06:04]
not very clear. And a better version of
[06:06]
this is saying use clear headers in each
[06:08]
section. Bold the key takeaways. Add a
[06:10]
summary at the top. We're very explicit
[06:12]
about what we want. The AI is likely
[06:14]
going to achieve that because it's very
[06:15]
good at following instructions. So
[06:17]
that's tip four. Ask for the formats
[06:18]
that you want instead of saying what you
[06:20]
don't want. Now on to tip five, which is
[06:22]
being direct about actions. What does
[06:24]
this mean? Well, often times when you
[06:25]
want AI to do something for you, we need
[06:27]
to avoid suggestive language. So instead
[06:30]
of saying suggest, think about,
[06:32]
consider, you need to be clear and
[06:34]
saying I want you to change this now,
[06:35]
edit this thing, or make whatever. by
[06:37]
being explicit about it taking action,
[06:40]
it will then take the action. But
[06:41]
oftentimes the AI can default to being
[06:44]
more oriented towards not taking action
[06:46]
to make to ensure it's not breaking
[06:48]
anything or doing anything it shouldn't.
[06:49]
So if you say, "Hey, I want you to
[06:51]
suggest this." It will suggest
[06:52]
something. It won't change anything. If
[06:54]
you want it to change something, then
[06:55]
you explicitly state that you want it to
[06:56]
change it. So be clear on the action
[06:59]
verbs that you're using with the AI. And
[07:01]
here are just some examples. So on the
[07:02]
left hand side we have a weak prompt
[07:04]
which we're stating can you suggest some
[07:06]
changes to improve this function. The
[07:07]
reason this is weak is we're stating
[07:09]
suggest. We're being more careful with
[07:11]
our language which means the AI will be
[07:13]
careful with its actions. It's not going
[07:14]
to do anything. It's only going to
[07:15]
suggest what we need to do to the
[07:17]
function. This is in relation to code. A
[07:19]
more explicit way of doing this that's
[07:21]
more actionoriented is saying change
[07:23]
this function to improve its
[07:24]
performance. That's when the AI goes off
[07:26]
and does the action for you. Another
[07:27]
example that's unrelated to code is
[07:29]
proposal prompts. So a weak version of
[07:31]
this is what do you think about this
[07:33]
proposal? This is just looking for
[07:34]
feedback, no actions. If you want an
[07:36]
action to occur from the AI, we would
[07:38]
say something that's more strong, which
[07:40]
is edit this proposal to make the
[07:41]
benefits clearer and add a call to
[07:43]
action at the end. This means the AI is
[07:45]
going to edit. It's going to do a thing
[07:46]
because there is an action verb inside
[07:48]
the prompt. So be clear about actions.
[07:50]
If you wanted to suggest something or
[07:52]
act acts on something, you need to be
[07:54]
clear about that. Now, on to tip six,
[07:55]
where Claude is actually extremely good
[07:57]
at research. It's gotten much better
[07:59]
over the last iteration of versions that
[08:01]
that Anthropics released with Opus 4.5.
[08:03]
Now, let me show you a weak and good
[08:05]
example. So, a weak example here for
[08:06]
research in relation to AI is saying,
[08:08]
"Research my competitors." This is not
[08:10]
good because it's vague and it's not
[08:11]
clear on what you want. A better version
[08:13]
of this is research my top three
[08:15]
competitors in the home services
[08:17]
industry. For each one, I want you to
[08:18]
find their pricing, their main services,
[08:20]
and their customer reviews. Compare them
[08:22]
to my business and tell me where I have
[08:23]
an advantage. So here we're explicit
[08:25]
about the type of research we wanted to
[08:27]
do, what we wanted to do with the
[08:28]
research after it's gotten back and a
[08:30]
series of other things. So this is what
[08:31]
a good research prompt looks like. And
[08:33]
something that's really interesting
[08:34]
about the guide that was provided for
[08:36]
anthropic is they actually provided a
[08:38]
metaprompt. So this specific prompt
[08:39]
right here can be applied to almost any
[08:41]
research use case that you have and it's
[08:44]
going to improve it. I've been using it
[08:45]
for the last couple of days and it's
[08:46]
night and day in the quality of the
[08:48]
research that I'm getting getting back
[08:49]
from the AI by using this prompt. So, I
[08:50]
highly recommend just copying and
[08:52]
pasting this from their guide and using
[08:53]
it inside your research. Now, why is
[08:55]
this so good? Well, there's a few things
[08:56]
as to why this prompt is so useful.
[08:58]
First off, we're asking the AI to
[08:59]
research in a structured way. So, be
[09:01]
more formal in your process. But also,
[09:03]
we're asking as you gather data, I want
[09:05]
you to develop competing hypotheses. So,
[09:08]
multiple opinions against what's
[09:10]
happening here. In addition to that,
[09:11]
we're also saying, can you uh track your
[09:13]
confidence levels over time and adjust
[09:15]
them as you learn new things? because
[09:17]
you're going to regularly self-critique
[09:19]
your confidence levels as well as the
[09:21]
hypotheses that you build over your
[09:22]
research process. And then we're also
[09:24]
noting that the research needs to be
[09:26]
broken down from a complex ask to more
[09:29]
manageable asks that you then
[09:30]
consolidate over time. So there's a lot
[09:32]
of really interesting information in
[09:34]
this prompt as to why it's useful, but I
[09:35]
highly recommend just using this for
[09:37]
your research going forward because
[09:38]
you're going to get much better
[09:39]
responses back for things that truly
[09:40]
matter. And then our final tip is
[09:43]
Claude's really good at making documents
[09:44]
specifically because it uses a claude
[09:46]
skill or variety of claude skills that
[09:48]
allows it to create documents that are
[09:49]
more compelling. So this is going to be
[09:51]
in the relation to presentations,
[09:53]
animations inside of presentations
[09:54]
and/or landing pages like this or any
[09:57]
type of visual documents that's relating
[09:58]
to a PDF, an Excel sheet, etc. So it's
[10:02]
generally better at creating these types
[10:04]
of documents and they tend to be
[10:05]
formatted more effectively instead of
[10:07]
what we used to get back in the day with
[10:08]
other models. So what are some examples
[10:10]
of this? Well, the first one here is a
[10:11]
presentation. So a bad prompt is make me
[10:14]
a presentation. A better prompt, a
[10:15]
stronger one is create me a professional
[10:17]
presentation on topic. You fill in the
[10:19]
topic, including thoughtful design
[10:20]
elements, visual hierarchy, and engaging
[10:22]
animations when appropriate. So we're
[10:24]
asking specifically about certain design
[10:26]
elements and asking it to think hard
[10:28]
about those elements. So when it
[10:29]
provides us back a presentation, it's
[10:30]
more compelling. So it's a better
[10:31]
prompt. Another example here is a
[10:34]
report. So instead of saying, "Write me
[10:36]
a report." Bad prompt. A better prompt
[10:38]
is create a monthly report for my team.
[10:40]
Include a summary at the top, sections
[10:42]
for each department, charts showing our
[10:44]
progress, action items for the ne for
[10:46]
the next month, and use clean formatting
[10:48]
that's easy to scan. So, we're asking
[10:50]
for a lot of things here. We're asking
[10:51]
explicitly what should be included in
[10:53]
the prompt. We're being very uh specific
[10:55]
on the inclusion. We're also being very
[10:57]
explicit on the formatting as well
[10:58]
because we want be people to be able to
[11:00]
scan this instead of having to read it
[11:02]
or struggle to read through it. And
[11:03]
that's our seventh tip. And as a quick
[11:05]
recap, here are just seven rules that I
[11:07]
pulled from Anthropic themselves on how
[11:09]
you can get most the most out of their
[11:10]
models. So the first one here is about
[11:13]
being clear. You want to be clear with
[11:14]
the model because it's very good at
[11:15]
following instructions. So be clear in
[11:17]
your prompt. The second thing is always
[11:19]
provide your intent. Explain why as to
[11:21]
why you're asking the AI to do this
[11:22]
thing for you. It can infer more from
[11:24]
that and give you a higher quality
[11:25]
response. After that, you should give it
[11:27]
high quality examples if you decide to
[11:29]
give it examples because the AI is going
[11:31]
to follow them to the tea. that includes
[11:33]
the prompt itself of how you're writing
[11:35]
it. So, be very careful on the prompt
[11:37]
that you're drafting. After that, if you
[11:39]
care about formatting, you want to be
[11:41]
explicit about that and ask for certain
[11:42]
types of formats from the AI. And that
[11:44]
means we're asking the AI what to do
[11:45]
instead of what not to do. After this,
[11:47]
we have the importance of being direct
[11:49]
with the AI. So, if we want it to take
[11:51]
action, we need to be clear on it taking
[11:53]
that action. If we wanted to suggest
[11:54]
something, we need to be clear on that.
[11:56]
Also, it's very good at research. And
[11:58]
that metaprompt that I showed you, you
[11:59]
should use that probably for good on all
[12:01]
your research going forward with the
[12:02]
specific model because it's night and
[12:03]
day compared to what I've gotten in the
[12:05]
past. And then finally, it's very good
[12:06]
at creating documents and it's good at
[12:08]
following different types of formatting
[12:10]
um for generalized reports, Excel
[12:12]
sheets, presentations, etc. And that's
[12:14]
it. So if you enjoyed this, reshare with
[12:15]
your friends. And also, as a reminder,
[12:17]
two things. First off, below is a 30-day
[12:19]
AI insight series, completely free.
[12:21]
You'll get 30 insights in your inbox how
[12:22]
you can apply AI to your business and
[12:24]
your work. Second thing is if you'd like
[12:25]
to work with me, there's a variety of
[12:27]
ways you can do that either through a
[12:28]
private AI community or directly
[12:30]
one-on-one. You can check that out below
[12:31]
as well. Okay, so now we know that cloud
[12:34]
can create great documents. Anthropic
[12:36]
even says so. But here's the thing. It
[12:38]
still can't match your formatting, your
[12:40]
fonts, your colors, your style. You
[12:42]
still have to copy the content and fix
[12:44]
it by hand in the document. I found one
[12:46]
cloud feature that fixes this. Now,
[12:48]
every document that comes back to me is
[12:50]
exactly how I want it. Let me show you
[12:52]
right here. Go ahead, click that video
[12:54]
right there.