1

00:00:00,003 --> 00:00:02,003

2
00:00:02,003 --> 00:00:04,004
but they're also unpredictable.

3
00:00:04,004 --> 00:00:07,005
Without guidance, they might choose the wrong language,

4
00:00:07,005 --> 00:00:11,007
switch frameworks or ignore your preferences entirely.

5
00:00:11,007 --> 00:00:13,009
System prompts act as guardrails,

6
00:00:13,009 --> 00:00:16,001
helping to align your assistant

7
00:00:16,001 --> 00:00:19,000
with your tools, styles, and goals.

8
00:00:19,000 --> 00:00:21,000
You can often set preferences

9
00:00:21,000 --> 00:00:23,002
globally across your environment

10
00:00:23,002 --> 00:00:25,005
so they apply to every project

11
00:00:25,005 --> 00:00:29,007
and/or define them locally for individual projects.

12
00:00:29,007 --> 00:00:31,003
That gives you a flexible way

13
00:00:31,003 --> 00:00:33,005
to keep your assistant grounded

14
00:00:33,005 --> 00:00:37,002
without having to repeat yourself every time you code.

15
00:00:37,002 --> 00:00:39,009
So both approaches help you maintain consistency
16
00:00:39,009 --> 00:00:41,006
when running your prompts,

17
00:00:41,006 --> 00:00:43,007
and each application will have

18
00:00:43,007 --> 00:00:45,009
a different place where you can do this.

19
00:00:45,009 --> 00:00:48,006
Whether you're new to coding or have some experience,

20
00:00:48,006 --> 00:00:50,008
it's a good idea to test out your prompts

21
00:00:50,008 --> 00:00:55,001
with a tool like the Claude Console or ChatGPT.

22
00:00:55,001 --> 00:00:57,003
You want to describe what you want to build

23
00:00:57,003 --> 00:01:00,009
even in plain language, and they can help you shape

24
00:01:00,009 --> 00:01:04,008
a more effective starting point for a system prompt.

25
00:01:04,008 --> 00:01:06,007
Just ask the AI to help you design

26
00:01:06,007 --> 00:01:08,006
the structure of your project

27
00:01:08,006 --> 00:01:10,008
and suggest a set of specs.

28
00:01:10,008 --> 00:01:13,002
One useful prompt might be something like,

29
00:01:13,002 --> 00:01:15,006
"I want to build an interactive dashboard

30
00:01:15,006 --> 00:01:18,001
"for exploring city level data.

31
00:01:18,001 --> 00:01:20,000
"Please ask me any questions you need

32
00:01:20,000 --> 00:01:22,007
"to fully understand what I want to build,

33
00:01:22,007 --> 00:01:25,002
"and then help me write a system prompt

34
00:01:25,002 --> 00:01:27,007
"that will guide the AI to consistently use

35
00:01:27,007 --> 00:01:30,006
"the right tools, frameworks, and coding styles

36
00:01:30,006 --> 00:01:32,004
"for this kind of project."

37
00:01:32,004 --> 00:01:36,000
This gives the AI a chance to gather important context first

38
00:01:36,000 --> 00:01:39,005
and helps you avoid having to correct assumptions later.

39
00:01:39,005 --> 00:01:40,008
From there, the system prompt

40
00:01:40,008 --> 00:01:43,000
you create together with your chat bots

41
00:01:43,000 --> 00:01:46,001
can serve as a foundation for your vibe coding session.

42
00:01:46,001 --> 00:01:47,006
Even if you're just starting out

43
00:01:47,006 --> 00:01:49,008
and don't have strong preferences,

44
00:01:49,008 --> 00:01:51,007
it still helps to set the direction.

45
00:01:51,007 --> 00:01:54,004
You can also use the Anthropic console tools.

46
00:01:54,004 --> 00:01:57,003
If you have an idea or a decent prompt,

47
00:01:57,003 --> 00:02:01,003
want to formalize it into a better system prompt structure.

48
00:02:01,003 --> 00:02:03,004
This tool gives you an option for models

49
00:02:03,004 --> 00:02:07,000
that have reasoning or thinking options.

50
00:02:07,000 --> 00:02:09,006
You can also test how the Claude model

51
00:02:09,006 --> 00:02:12,008
reacts to your prompt using the Evaluate

52
00:02:12,008 --> 00:02:14,009
and Generate Test Case buttons.

53
00:02:14,009 --> 00:02:17,009
These allow you to simulate multiple interactions

54
00:02:17,009 --> 00:02:22,004
and check for consistency and quality in the AI responses.

55
00:02:22,004 --> 00:02:25,003
Although the Anthropic console is specific to Claude,

56
00:02:25,003 --> 00:02:28,001
the resulting prompts will often work well

57
00:02:28,001 --> 00:02:29,004
with other platforms.

58
00:02:29,004 --> 00:02:31,005
That prompt will act as a blueprint,

59
00:02:31,005 --> 00:02:34,007
giving the assistant what it needs to stay consistent
60
00:02:34,007 --> 00:02:37,006
while still letting you explore as you build.

61
00:02:37,006 --> 00:02:38,009
For experienced users,

62
00:02:38,009 --> 00:02:41,001
the AI is going to make some assumptions

63
00:02:41,001 --> 00:02:43,008
based on how you ask a question,

64
00:02:43,008 --> 00:02:47,004
and that can sometimes lead to unpredictable results.

65
00:02:47,004 --> 00:02:50,002
You might get code in a language you're not using

66
00:02:50,002 --> 00:02:52,007
or see tools that are introduced

67
00:02:52,007 --> 00:02:54,003
that don't fit your project.

68
00:02:54,003 --> 00:02:56,007
So for example, you can say something as simple as,

69
00:02:56,007 --> 00:02:59,008
"Use JavaScript, not TypeScript or Python,"

70
00:02:59,008 --> 00:03:02,003
"Use react for UI components,"

71
00:03:02,003 --> 00:03:05,000
"Style everything with Tailwind CSS,"

72
00:03:05,000 --> 00:03:08,006
"Avoid using external libraries unless specified."

73
00:03:08,006 --> 00:03:11,007
So that will help the AI's do the right thing

74
00:03:11,007 --> 00:03:15,003
and continue to work within preset parameters
75
00:03:15,003 --> 00:03:17,001
that developers will understand.

76
00:03:17,001 --> 00:03:18,009
Now, popular frameworks and tools

77
00:03:18,009 --> 00:03:22,001
are often a better choice when working with AI

78
00:03:22,001 --> 00:03:25,009
because the models have more examples of those.

79
00:03:25,009 --> 00:03:28,005
For instance, React is the most widely used

80
00:03:28,005 --> 00:03:31,002
front-end framework for building interactive content.

81
00:03:31,002 --> 00:03:33,005
But if you know your team is working with Vue,

82
00:03:33,005 --> 00:03:35,009
that's also important to include.

83
00:03:35,009 --> 00:03:39,003
Similarly, Tailwind is a common CSS framework,

84
00:03:39,003 --> 00:03:41,003
and JavaScript libraries like D3

85
00:03:41,003 --> 00:03:43,008
are well suited for visualizing data

86
00:03:43,008 --> 00:03:48,000
as long as you want to build interactive web components.

87
00:03:48,000 --> 00:03:51,007
If you're working with other data analysis or automation,

88
00:03:51,007 --> 00:03:53,009
Python might be a better fit.

89
00:03:53,009 --> 00:03:57,001
Choosing a language that the AI will understand well

90
00:03:57,001 --> 00:03:59,008
and potentially something that you know how to read

91
00:03:59,008 --> 00:04:02,003
will usually lead to better results.

92
00:04:02,003 --> 00:04:04,003
Now with vibe coding, clarity upfront

93
00:04:04,003 --> 00:04:07,005
is going to lead to better outcomes later.

94
00:04:07,005 --> 00:04:09,003
System prompts let you control the tools

95
00:04:09,003 --> 00:04:12,003
and the tone of your projects from the start

96
00:04:12,003 --> 00:04:14,007
so that the AI is really helping you build,

97
00:04:14,007 --> 00:04:18,000
and you don't have to backtrack as much in regular projects.