1
00:00:00,000 --> 00:00:03,760
So I- I was sitting at the bench
yesterday, trying to push this tiny,

2
00:00:03,824 --> 00:00:05,920
tiny little CSS tweak, right?

3
00:00:05,968 --> 00:00:11,280
Just a, a margin fix, and suddenly my
terminal goes absolutely wild.

4
00:00:11,413 --> 00:00:12,960
It's like, uh...

5
00:00:12,987 --> 00:00:17,440
it's like Claude decided to spin up a
whole integration suite and- and run a

6
00:00:17,493 --> 00:00:19,440
security audit without me even asking.

7
00:00:19,760 --> 00:00:22,960
I was literally just sitting there
watching my token balance burn.

8
00:00:24,198 --> 00:00:27,797
Oh man, the classic, the classic rogue
agent spiral.

9
00:00:27,858 --> 00:00:31,878
But, you know, that is actually exactly
what Anthropic is putting a stop to with the

10
00:00:31,997 --> 00:00:34,677
new Claude Code update that dropped on
July nineteenth.

11
00:00:34,708 --> 00:00:35,508
Wait, really?

12
00:00:35,636 --> 00:00:36,868
July nineteenth?

13
00:00:37,108 --> 00:00:39,588
What- what did they actually change?

14
00:00:39,782 --> 00:00:46,023
Yeah, so in version 2.1.215, they
completely pulled back on that unprompted

15
00:00:46,063 --> 00:00:46,522
behavior.

16
00:00:46,583 --> 00:00:50,862
The- the big headline here is that Claude
no longer runs the /verify and

17
00:00:50,882 --> 00:00:52,503
/code-review skills on its own.

18
00:00:53,282 --> 00:00:54,742
Oh, thank goodness for that.

19
00:00:55,681 --> 00:00:59,221
Seriously, it was like working with a
super eager junior engineer,

20
00:00:59,421 --> 00:00:59,641
you know?

21
00:01:00,241 --> 00:01:03,882
Like, you ask them to fix one typo on a
button, and they, uh...

22
00:01:04,701 --> 00:01:09,582
they decide to independently trigger a
multi-stage integration test and a container

23
00:01:09,642 --> 00:01:11,361
vulnerability scan.

24
00:01:11,401 --> 00:01:14,761
You're just sitting there, tapping your
teacup, waiting for the cloud bill to

25
00:01:14,781 --> 00:01:15,222
arrive.

26
00:01:15,769 --> 00:01:16,109
Right!

27
00:01:16,189 --> 00:01:16,750
Exactly.

28
00:01:17,250 --> 00:01:22,769
You get a massive context-window bloat and
this huge execution lag for a- a one-line

29
00:01:22,829 --> 00:01:23,210
change.

30
00:01:23,510 --> 00:01:24,130
It's too much.

31
00:01:24,729 --> 00:01:26,670
But now, it's- it's entirely manual.

32
00:01:26,729 --> 00:01:28,470
The developer is back in the driver's
seat.

33
00:01:28,969 --> 00:01:33,770
If you want to test, you have to
explicitly type /verify to run your local test

34
00:01:33,809 --> 00:01:34,170
scripts.

35
00:01:34,629 --> 00:01:35,850
It won't just guess anymore.

36
00:01:36,054 --> 00:01:37,854
Which, I mean, it makes a lot of sense.

37
00:01:38,234 --> 00:01:38,494
But...

38
00:01:39,714 --> 00:01:44,635
okay, if we have to do it ourselves now,
how does the, the workflow actually change?

39
00:01:45,335 --> 00:01:46,914
Like, what's the playbook?

40
00:01:47,000 --> 00:01:49,560
Well, the power move now is using the
flags.

41
00:01:49,587 --> 00:01:54,360
So, if you're in an interactive session,
you can run /code-review --fix.

42
00:01:54,472 --> 00:01:59,160
That'll look at your local git diff, find
the issues, and actually apply the fixes

43
00:01:59,196 --> 00:02:00,520
straight to your working tree.

44
00:02:00,640 --> 00:02:01,120
Oh, nice.

45
00:02:01,659 --> 00:02:04,839
So it still does the heavy lifting, you
just have to give it the nod first.

46
00:02:04,917 --> 00:02:05,517
Precisely.

47
00:02:05,527 --> 00:02:10,517
And if you want to go, like, full
enterprise mode, there's /code-review ultra.

48
00:02:10,613 --> 00:02:15,557
That actually dispatches a whole fleet of
remote reviewer agents on Anthropic's

49
00:02:15,624 --> 00:02:18,677
cloud to do some deep, verified
bug-hunting.

50
00:02:18,730 --> 00:02:19,477
It's pretty wild.

51
00:02:20,558 --> 00:02:21,678
Remote fleet, eh?

52
00:02:21,958 --> 00:02:25,637
That sounds like a- a proper way to burn
through some serious compute,

53
00:02:25,878 --> 00:02:28,338
but at least now we're the ones pulling
the trigger.

54
00:02:29,258 --> 00:02:32,577
But, you know, this does mean we've got to
be a bit more disciplined.

55
00:02:32,657 --> 00:02:35,937
If we don't manually run /verify before we
push...

56
00:02:36,837 --> 00:02:38,877
well, that safety net is gone.

57
00:02:39,677 --> 00:02:42,138
We could easily deploy broken code if we
get lazy.

58
00:02:42,386 --> 00:02:43,006
Oh, totally.

59
00:02:43,046 --> 00:02:44,126
The training wheels are off.

60
00:02:44,607 --> 00:02:48,426
But hey, they did add some great
quality-of-life stuff in the background too.

61
00:02:48,526 --> 00:02:51,746
Like, long-running tool calls finally have
a progress heartbeat now,

62
00:02:52,167 --> 00:02:54,526
so the terminal doesn't just look frozen
while it's thinking.

63
00:02:54,723 --> 00:02:55,302
Ah, yes!

64
00:02:55,782 --> 00:02:59,342
The "is it dead or is it just
contemplating life" problem.

65
00:02:59,983 --> 00:03:01,063
That's a massive win.

66
00:03:01,762 --> 00:03:04,162
And hey, didn't they fix that Linux bug
too?

67
00:03:04,662 --> 00:03:08,463
The one where if you ran a broad pkill -f
inside the bash tool,

68
00:03:08,522 --> 00:03:08,882
it would...

69
00:03:09,562 --> 00:03:10,603
uh, what did it do?

70
00:03:10,962 --> 00:03:12,122
Target its own CLI?

71
00:03:12,806 --> 00:03:14,846
Yeah, it would literally commit suicide.

72
00:03:14,907 --> 00:03:18,626
It would kill Claude's own Node CLI runner
process.

73
00:03:18,686 --> 00:03:22,206
That is finally fixed in the v2.1.214
patches.

74
00:03:22,786 --> 00:03:24,446
No more accidental self-destruction.

75
00:03:25,240 --> 00:03:25,639
Brilliant.

76
00:03:26,219 --> 00:03:27,819
Well, it's good to have the reins back.

77
00:03:28,259 --> 00:03:30,539
I'm off to manually verify some CSS.

78
00:03:31,179 --> 00:03:31,899
Catch you later, mate.

79
00:03:32,583 --> 00:03:33,542
Sounds good, talk soon.