1
00:00:00,000 --> 00:00:01,194
Welcome to the show!

2
00:00:01,194 --> 00:00:02,868
I'm Lachlan Reed with James Turner,

3
00:00:02,868 --> 00:00:07,711
and mate, I wanna start with one tiny
little word that's gonna matter

4
00:00:07,711 --> 00:00:10,215
to a lot of devs: xhigh.

5
00:00:10,215 --> 00:00:16,367
Claude Code 2.

6
00:00:10,215 --> 00:00:16,367
1.111 adds an xhigh effort tier for Opus
4.

7
00:00:10,215 --> 00:00:16,367
7,

8
00:00:16,367 --> 00:00:19,323
and it sits right between high and max.

9
00:00:19,323 --> 00:00:25,096
Not "faster mode," not "genius mode" --
more like that middle gear where

10
00:00:25,096 --> 00:00:29,212
you get noticeably more reasoning than
high without copping the full token

11
00:00:29,212 --> 00:00:33,455
bill and latency hit of max.
The phrase that jumps out to me

12
00:00:33,455 --> 00:00:37,759
is "between high and max." Because that
means a lot of people were living

13
00:00:37,759 --> 00:00:41,323
in this weird binary: either high is good
enough,

14
00:00:41,323 --> 00:00:44,036
or you smash the red glass and go max.

15
00:00:44,036 --> 00:00:48,550
xhigh is basically admitting that "good
enough" and "go nuclear" was

16
00:00:48,550 --> 00:00:51,917
a clumsy choice for real work.

17
00:00:48,550 --> 00:00:51,917
Exactly.

18
00:00:51,917 --> 00:00:54,345
And if you've done enough coding with
these tools,

19
00:00:54,345 --> 00:00:55,545
you feel that gap.

20
00:00:55,545 --> 00:00:59,617
High is usually fine for routine edits --
clean up a component,

21
00:00:59,617 --> 00:01:01,692
rename a few functions, patch a test.

22
00:01:01,692 --> 00:01:05,287
But then you hit the grubby stuff:
architecture reviews,

23
00:01:05,287 --> 00:01:07,843
some cursed async race condition,

24
00:01:07,843 --> 00:01:12,255
or a refactor where every file is
attached to every other file like

25
00:01:07,843 --> 00:01:12,255
spaghetti

26
00:01:12,255 --> 00:01:13,105
in a lawn mower.

27
00:01:13,105 --> 00:01:17,096
That's where you want the model to slow
down and think harder...

28
00:01:17,096 --> 00:01:22,411
just not max-hard on every
prompt."Async race

29
00:01:17,096 --> 00:01:22,411
condition"

30
00:01:22,411 --> 00:01:24,036
is the exact phrase there.

31
00:01:24,036 --> 00:01:25,396
Because for something like that,

32
00:01:25,396 --> 00:01:28,174
one missed assumption and the whole
answer's junk.

33
00:01:28,174 --> 00:01:32,916
So I get the appeal of "just enough extra
brainpower." You're not asking

34
00:01:32,916 --> 00:01:34,073
for infinite thought.

35
00:01:34,073 --> 00:01:39,200
You're asking for the model to take the
problem SERIOUSLY.Yeah,

36
00:01:39,200 --> 00:01:42,398
give it another coffee,
not a full existential retreat.

37
00:01:42,398 --> 00:01:45,076
That's the practical charm.

38
00:01:45,076 --> 00:01:47,510
If I'm doing a routine UI tweak,

39
00:01:47,510 --> 00:01:48,710
I don't need max.

40
00:01:48,710 --> 00:01:53,422
But if I'm staring at a nasty refactor
and I want cleaner reasoning paths,

41
00:01:53,422 --> 00:01:55,893
xhigh sounds like the setting where I'd
go,

42
00:01:55,893 --> 00:02:00,019
"Alright, settle in, have a proper
think."But let me push on the other

43
00:02:00,019 --> 00:02:00,355
side.

44
00:02:00,355 --> 00:02:03,784
Does adding xhigh actually simplify the
decision?

45
00:02:03,784 --> 00:02:07,446
Or does it create one more knob for
developers to obsess over?

46
00:02:07,446 --> 00:02:10,804
Because devs love turning one dial into a
religion.

47
00:02:10,804 --> 00:02:13,199
Suddenly somebody's benchmarking low,

48
00:02:13,199 --> 00:02:18,262
normal, high, xhigh, max across ten
prompts and making a spreadsheet

49
00:02:18,262 --> 00:02:23,034
instead of shipping.That is
SUCH a developer problem.

50
00:02:23,034 --> 00:02:26,698
We'll spend 40 minutes tuning the wrench
before tightening the bolt.

51
00:02:26,698 --> 00:02:31,494
But I think this one earns its keep
because it maps to a real felt gap.

52
00:02:31,494 --> 00:02:35,171
It's not some mystery setting wedged in
for marketing.

53
00:02:35,171 --> 00:02:39,164
It answers a common moment: "I need more
than high,

54
00:02:39,164 --> 00:02:44,239
but max feels like overkill."And
"overkill" is the important word.

55
00:02:44,239 --> 00:02:49,078
Max can be worth it, sure, but if you're
reaching for it on every serious

56
00:02:49,078 --> 00:02:52,368
prompt, you're basically saying the
ladder has a missing rung.

57
00:02:52,368 --> 00:02:55,553
xhigh adds that rung.Right.

58
00:02:55,553 --> 00:02:59,150
And once a tool gives you that middle
ground,

59
00:02:59,150 --> 00:03:00,425
it changes your habits.

60
00:03:00,425 --> 00:03:02,658
You stop thinking in extremes.

61
00:03:02,658 --> 00:03:05,774
Not every hard problem needs the
sledgehammer.

62
00:03:05,774 --> 00:03:09,778
Sometimes you just need a slightly
sharper chisel.

63
00:03:05,774 --> 00:03:09,778
The part I actually

64
00:03:09,778 --> 00:03:12,988
love here is the redesigned /effort
command.

65
00:03:12,988 --> 00:03:16,822
If you run /effort with no arguments now,

66
00:03:16,822 --> 00:03:19,199
it opens an interactive arrow-key slider.

67
00:03:19,199 --> 00:03:23,193
You move across low, normal, high,

68
00:03:23,193 --> 00:03:26,311
xhigh, and max, then hit Enter to confirm.

69
00:03:26,311 --> 00:03:30,555
That sounds minor until you've spent
enough years in terminals to know

70
00:03:30,555 --> 00:03:36,236
the difference between "feature exists"
and "feature is usable under

71
00:03:30,555 --> 00:03:36,236
pressure."

72
00:03:36,236 --> 00:03:37,828
Ooh, that's a good distinction.

73
00:03:37,828 --> 00:03:42,084
Because a hidden config trick always
feels a bit like finding a spanner

74
00:03:42,084 --> 00:03:45,405
in the shed after you've already tried
fixing the bike with a spoon.

75
00:03:45,405 --> 00:03:48,357
If it's interactive -- arrow keys,

76
00:03:48,357 --> 00:03:52,018
live choice, press Enter -- suddenly it
feels like part of the workflow,

77
00:03:52,018 --> 00:03:56,799
not trivia from page 6 of the docs.

78
00:03:52,018 --> 00:03:56,799
Yes.

79
00:03:56,799 --> 00:03:59,592
"Arrow-key slider" is the thing I'd react
to.

80
00:03:59,592 --> 00:04:03,578
That turns effort into something you
adjust live while experimenting,

81
00:04:03,578 --> 00:04:07,087
not something you memorize as a flag and
forget exists.

82
00:04:07,087 --> 00:04:10,496
You can feel your way through it in the
terminal.

83
00:04:07,087 --> 00:04:10,496
And there are still

84
00:04:10,496 --> 00:04:11,911
the practical entry points too.

85
00:04:11,911 --> 00:04:15,425
You can launch with claude --effort xhigh,

86
00:04:15,425 --> 00:04:18,536
which is dead handy if you already know
what session you want.

87
00:04:18,536 --> 00:04:22,526
Or pick it in the model picker UI if you
want that visible,

88
00:04:22,526 --> 00:04:23,637
repeatable choice.

89
00:04:23,637 --> 00:04:26,751
So it's not just one path -- it's
terminal,

90
00:04:26,751 --> 00:04:32,840
launch flag, and UI all lining
up."claude --effort xhigh"

91
00:04:32,840 --> 00:04:35,656
is exactly the kind of thing teams can
standardize around.

92
00:04:35,656 --> 00:04:39,010
But I think the interactive piece matters
more for discovery.

93
00:04:39,010 --> 00:04:43,444
I've had so many CLI moments where I
technically knew a feature existed

94
00:04:43,444 --> 00:04:47,922
because I read the docs, but it didn't
become REAL until I used it and

95
00:04:47,922 --> 00:04:50,795
felt the shape of it.Oh,

96
00:04:50,795 --> 00:04:51,511
absolutely.

97
00:04:51,511 --> 00:04:53,974
I've wrestled enough CLIs to know this
pain.

98
00:04:53,974 --> 00:04:56,530
You read a command, nod wisely,

99
00:04:56,530 --> 00:04:59,161
then three days later you remember none
of it.

100
00:04:59,161 --> 00:05:02,915
But when a terminal lets you poke at it
-- move left,

101
00:05:02,915 --> 00:05:05,948
move right,
see the options -- your brain goes,

102
00:05:05,948 --> 00:05:10,318
"Ahh, got it." It's the difference
between reading a map and actually

103
00:05:10,318 --> 00:05:13,358
riding the trail.

104
00:05:10,318 --> 00:05:13,358
And for experimentation,

105
00:05:13,358 --> 00:05:14,638
that's huge.

106
00:05:14,638 --> 00:05:17,985
Because effort level is not a one-time
preference.

107
00:05:17,985 --> 00:05:19,824
It changes with the task.

108
00:05:19,824 --> 00:05:23,491
You might start at high,
bump to xhigh when the prompt gets hairy,

109
00:05:23,491 --> 00:05:24,450
then go back down.

110
00:05:24,450 --> 00:05:29,200
A live control surface fits that reality
better than editing config files

111
00:05:29,200 --> 00:05:33,304
every time.That's the
phrase: control surface.

112
00:05:33,304 --> 00:05:34,504
Not just a flag.

113
00:05:34,504 --> 00:05:39,286
Once the model picker UI and the slash
command both expose the same choice,

114
00:05:39,286 --> 00:05:43,557
you're telling users, "This is a knob
you're supposed to use." Not a

115
00:05:43,557 --> 00:05:47,476
secret handshake.

116
00:05:43,557 --> 00:05:47,476
And when a feature stops feeling like a
secret,

117
00:05:47,476 --> 00:05:48,754
adoption goes up.

118
00:05:48,754 --> 00:05:52,462
Not because the docs got better -- though
docs matter -- but because

119
00:05:52,462 --> 00:05:56,025
the interface finally admits what people
are actually trying to do in

120
00:05:56,025 --> 00:05:56,962
the middle of work.