WEBVTT

00:00:00.080 --> 00:00:03.760
Now, when working with AI agents,
a key part of your entire

00:00:03.800 --> 00:00:07.490
work is writing good prompts
and providing the right

00:00:07.540 --> 00:00:11.420
context. Now,
in simple projects like this,

00:00:11.500 --> 00:00:15.400
file, the AI, as you saw,
will typically do a good job

00:00:15.440 --> 00:00:18.800
at finding the relevant files
and gathering the relevant

00:00:19.100 --> 00:00:23.000
context. And, in general, over time,
AI models and these tools

00:00:23.040 --> 00:00:26.710
get better,
so they tend to get the context they

00:00:26.740 --> 00:00:30.060
need. But context is a key word here

00:00:30.360 --> 00:00:34.040
because when working with these agents,
you should

00:00:34.100 --> 00:00:37.880
try to craft good prompts.
Prompt engineering is important

00:00:37.920 --> 00:00:40.780
because good prompts yield better results.

00:00:40.790 --> 00:00:44.780
And that starts with being specific,
just as you would try to

00:00:44.840 --> 00:00:48.060
be specific if you gave a coworker a

00:00:48.120 --> 00:00:52.060
task.
It's also about providing useful context,

00:00:52.080 --> 00:00:55.200
though,
to steer the AI in the right direction.

00:00:55.220 --> 00:00:59.140
And that's even useful if you think the AI

00:00:59.240 --> 00:01:02.940
will do a good job at finding the right
context because explicitly

00:01:03.000 --> 00:01:06.860
giving the AI the context you know it
needs is, of course, better

00:01:06.880 --> 00:01:09.920
than hoping that it will discover it.

00:01:09.960 --> 00:01:13.880
It's also potentially about providing
useful examples depending on the

00:01:13.940 --> 00:01:17.720
task,
and an example could be an example you put

00:01:17.740 --> 00:01:21.240
But it could also be a pointer at some
other file where you already have an

00:01:21.260 --> 00:01:24.620
example. And you wanna split complex

00:01:24.680 --> 00:01:28.520
tasks.
You don't wanna have one prompt for

00:01:28.620 --> 00:01:30.940
your entire super complex project.

00:01:30.960 --> 00:01:34.660
Instead,
you wanna build it feature by feature with

00:01:34.880 --> 00:01:38.540
sessions where every session
is about one feature or one

00:01:38.680 --> 00:01:42.620
part of a feature because you will get
better results if you focus

00:01:42.680 --> 00:01:46.120
on specific tasks instead of asking for
everything in one

00:01:46.160 --> 00:01:50.100
go. And it's also about iterating,
iterating to

00:01:50.140 --> 00:01:52.760
improve the result because it's a chat.

00:01:52.900 --> 00:01:56.780
You can ask follow-up questions
or ask for follow-up

00:01:57.120 --> 00:02:00.680
edits. But don't get lost in iteration

00:02:00.740 --> 00:02:04.070
hell. If you see
that you're going nowhere and it fails

00:02:04.080 --> 00:02:07.880
consistently to fix a certain problem,
fix it yourself,

00:02:08.000 --> 00:02:11.109
start a new session.
Don't keep on iterating forever.

00:02:11.140 --> 00:02:13.890
You'll just burn tokens, which you,
of course, are

00:02:13.940 --> 00:02:15.220
paying.