1
00:00:00,100 --> 00:00:07,410
So it's late 2025 and to be 
honest I've stopped reading the 

2
00:00:07,420 --> 00:00:08,029
change logs. 
I. 

3
00:00:08,039 --> 00:00:09,910
Don't blame you at all, it's 
impossible to keep. 

4
00:00:09,920 --> 00:00:11,370
Up. 
It feels like every 5 minutes 

5
00:00:11,380 --> 00:00:15,470
there's a new AI coding tool, a 
new plug in, and it all promises

6
00:00:15,480 --> 00:00:19,710
to revolutionize my workflow. 
I'm just feeling this massive 

7
00:00:19,760 --> 00:00:20,610
update. 
Fatigue. 

8
00:00:20,620 --> 00:00:23,110
It's the paradox of choice. 
We're drowning an option? 

9
00:00:23,120 --> 00:00:26,400
Exactly. 
So welcome everyone to the 

10
00:00:26,410 --> 00:00:28,020
architecture corner. 
Today. 

11
00:00:28,030 --> 00:00:30,290
We are just stripping all that 
back. 

12
00:00:30,300 --> 00:00:31,700
We're not looking at this shiny 
new stuff. 

13
00:00:31,710 --> 00:00:34,920
We're going back to the vanilla 
experience evolved, looking at 

14
00:00:34,930 --> 00:00:38,120
GitHub Copilot and VS Code and 
figuring out how to move from 

15
00:00:38,130 --> 00:00:42,160
just, I don't know, generating 
code to actual architectural 

16
00:00:42,170 --> 00:00:44,300
planning. 
The goal here is precision. 

17
00:00:44,310 --> 00:00:47,020
It's about making the tools 
behave, not letting them 

18
00:00:47,030 --> 00:00:49,940
overwhelm you. 
And I think the biggest shift is

19
00:00:49,950 --> 00:00:52,300
how we're starting to use custom
instructions. 

20
00:00:52,370 --> 00:00:54,480
Right, this is so much more than
just the prompt. 

21
00:00:54,490 --> 00:00:56,360
The prompt is what one the 
instructions file. 

22
00:00:56,370 --> 00:00:58,990
That's how you want it done? 
The style, the principles? 

23
00:00:59,060 --> 00:01:00,790
Yeah, the soul of the project. 
Almost. 

24
00:01:01,180 --> 00:01:03,770
People make a huge mistake here.
They just dump everything into 

25
00:01:03,780 --> 00:01:07,650
one giant markdown file. 
And that's terrible practice. 

26
00:01:07,720 --> 00:01:09,510
It's awful. 
It triggers something called 

27
00:01:09,520 --> 00:01:10,930
context rot. 
OK. 

28
00:01:10,940 --> 00:01:12,910
So that's tied to that lost in 
the middle problem. 

29
00:01:12,920 --> 00:01:17,130
With LM precisely, the model is 
great at remembering the very 

30
00:01:17,140 --> 00:01:19,490
beginning and the very end of 
your instructions. 

31
00:01:19,500 --> 00:01:22,370
But the stuff in the middle. 
It gets forgotten completely. 

32
00:01:22,440 --> 00:01:25,530
Yeah, If you bury your testing 
protocols or your naming 

33
00:01:25,540 --> 00:01:29,940
conventions deep inside a 2000 
line file, the model basically 

34
00:01:29,950 --> 00:01:32,000
hallucinates that they don't 
even exist. 

35
00:01:32,050 --> 00:01:34,600
So the fix have to be 
granularity. 

36
00:01:34,650 --> 00:01:36,600
You have to break it. 
U you have to modularize. 

37
00:01:36,610 --> 00:01:40,140
Don't use one mega file. 
You create a tiny file just for 

38
00:01:40,150 --> 00:01:43,520
testing, telling it exactly 
which frameworks and mocks to 

39
00:01:43,530 --> 00:01:44,740
use. 
You have another one for 

40
00:01:44,750 --> 00:01:46,620
security. 
So you're only injecting the 

41
00:01:46,630 --> 00:01:48,960
context that's relevant for that
specific. 

42
00:01:48,970 --> 00:01:51,140
Task. 
It's like a dependency injection

43
00:01:51,150 --> 00:01:53,900
for your AI's brain. 
That's a perfect analogy. 

44
00:01:53,910 --> 00:01:57,060
And that way of thinking, it 
owes a lot to Mario Bittencourt.

45
00:01:57,130 --> 00:02:00,230
He's really been pushing this 
idea of practical application, 

46
00:02:00,240 --> 00:02:03,070
not just hype. 
Which brings us to MCP. 

47
00:02:03,360 --> 00:02:06,290
The model context protocol. 
Right, this is the bridge to the

48
00:02:06,300 --> 00:02:08,449
outside world. 
Because the model itself is sort

49
00:02:08,460 --> 00:02:10,370
of frozen in time, right? 
Yeah. 

50
00:02:10,460 --> 00:02:13,870
If I ask for an API spec, it's 
gonna give me whatever it 

51
00:02:13,880 --> 00:02:16,650
learned during training, which 
could be two years out of date. 

52
00:02:16,700 --> 00:02:20,220
Exactly, but with MCP the AI can
use tools. 

53
00:02:20,620 --> 00:02:23,110
Bittencourt shows a great 
example with a fetch tool. 

54
00:02:23,200 --> 00:02:26,810
Instead of guessing an API, the 
model literally goes out, grabs 

55
00:02:26,820 --> 00:02:30,640
the live open API, ARC version 
3, point 2.0, whatever is 

56
00:02:30,650 --> 00:02:34,100
current and it codes against. 
That, but I see a risk there. 

57
00:02:34,170 --> 00:02:37,840
I mean, what if we start adding 
an MCP server for AW, another to

58
00:02:37,850 --> 00:02:40,870
the database, another for docs? 
You get tool overload, the model

59
00:02:40,880 --> 00:02:43,560
starts getting confused about 
which tool to pick, and even 

60
00:02:43,570 --> 00:02:45,980
worse, you open up a huge 
security vector. 

61
00:02:45,990 --> 00:02:49,220
Prompt injection. 
If you connect your AI to a 

62
00:02:49,230 --> 00:02:52,120
malicious MCP server, you've 
basically given an outsider a 

63
00:02:52,130 --> 00:02:55,880
shell you could be leaking data.
So the rule is use MCP, but you 

64
00:02:55,890 --> 00:02:58,040
absolutely have to audit the 
servers you install. 

65
00:02:58,430 --> 00:03:01,060
That is a fair warning. 
OK, so let's pivot to skills. 

66
00:03:01,070 --> 00:03:03,100
Yeah. 
How is a skill different from 

67
00:03:03,110 --> 00:03:04,390
the instructions we were just 
talking? 

68
00:03:04,400 --> 00:03:06,860
About it's a great question. 
Instructions are context, 

69
00:03:06,870 --> 00:03:09,120
they're facts. 
Skills are workflows. 

70
00:03:09,130 --> 00:03:12,380
They are checklists, standard 
operating procedures. 

71
00:03:12,510 --> 00:03:15,520
So it's not just write code, 
it's follow these exact. 

72
00:03:15,530 --> 00:03:19,700
Steps exactly right. 
A skill forces the AI to execute

73
00:03:19,710 --> 00:03:23,760
a specific multi step process 
like a web performance audit. 

74
00:03:24,110 --> 00:03:27,120
The skill doesn't let it guess, 
it has to, you know, one drop. 

75
00:03:27,130 --> 00:03:32,200
Run a trace 2. 
Analyze Core Web Vitals 3. 

76
00:03:32,500 --> 00:03:35,750
Check accessibility. 
It enforces a standard. 

77
00:03:35,760 --> 00:03:37,980
That creates consistency, which 
is huge. 

78
00:03:38,460 --> 00:03:40,990
But the feature that really 
sounds like a game changer to me

79
00:03:41,000 --> 00:03:43,690
is Background Agents. 
Ah, this is where we solved the 

80
00:03:43,700 --> 00:03:46,830
human multitasking problem. 
Because we are just awful at 

81
00:03:46,840 --> 00:03:48,450
contact switching. 
Terrible. 

82
00:03:48,560 --> 00:03:51,630
But an agent isn't. 
You can kick off a big task, say

83
00:03:51,640 --> 00:03:55,330
refactor this entire legacy 
module and the agent just spins 

84
00:03:55,340 --> 00:03:57,910
it up in a local git work. 
Tree, work tree. 

85
00:03:57,920 --> 00:04:00,360
OK, that's the key detail. 
So it's running asynchronously, 

86
00:04:00,440 --> 00:04:03,950
a totally parallel branch. 
It's not messing with my 

87
00:04:03,960 --> 00:04:05,100
current. 
Work exactly. 

88
00:04:05,110 --> 00:04:07,880
You could have one agent 
refactoring the back end schemas

89
00:04:07,890 --> 00:04:10,330
while another one is updating 
the front end components to 

90
00:04:10,340 --> 00:04:12,570
match, all while you were 
working on something else 

91
00:04:12,580 --> 00:04:14,270
entirely. 
When they're done, you just 

92
00:04:14,280 --> 00:04:16,390
merge the work. 
Trees like parallel processing 

93
00:04:16,399 --> 00:04:19,029
for development work it. 
Is, but you have to remember 

94
00:04:19,040 --> 00:04:21,010
Bittencourt's core philosophy 
here. 

95
00:04:21,100 --> 00:04:23,950
In specs and adapt. 
Yeah, don't just throw out your 

96
00:04:23,960 --> 00:04:26,390
entire workflow because we said 
agents sound cool. 

97
00:04:26,920 --> 00:04:29,830
Test it. 
If modular instruction files 

98
00:04:29,840 --> 00:04:32,880
actually stop the AI from 
forgetting your testing rules, 

99
00:04:33,170 --> 00:04:35,800
then adopt that. 
If a background agent genuinely 

100
00:04:35,810 --> 00:04:39,860
saves you 20 minutes of context 
switching a day, use it, but 

101
00:04:39,870 --> 00:04:42,310
solve a real problem, don't just
chase the tool. 

102
00:04:42,390 --> 00:04:45,060
That's the only way we're going 
to survive this fatigue. 

103
00:04:45,790 --> 00:04:48,260
If you want to see the diagrams 
and the config examples we 

104
00:04:48,270 --> 00:04:50,640
talked about, please do check 
out the description. 

105
00:04:50,690 --> 00:04:52,450
It's worth a. 
Look and for this kind of deep 

106
00:04:52,460 --> 00:04:55,780
analysis every week. 
Subscribe for free to the 

107
00:04:55,790 --> 00:04:57,080
Architecture Corner newsletter 
at 

108
00:04:57,090 --> 00:05:00,920
architecturecorner.substack.com.
It is the best way to stay 

109
00:05:00,930 --> 00:05:03,600
current without burning out. 
Thanks for listening. 

110
00:05:03,890 --> 00:05:05,720
Now go clean up those markdown 
files.