1
00:00:00,079 --> 00:00:04,559
Quick shout out before we dive in, this
episode is brought to you by Jellypod AI.

2
00:00:05,159 --> 00:00:08,179
Turn your written content into audio
effortlessly.

3
00:00:08,469 --> 00:00:09,370
Yeah, check them out!

4
00:00:09,849 --> 00:00:13,609
Now, okay, Lachlan, we, we, we need to
talk about something that's been giving

5
00:00:13,630 --> 00:00:16,609
engineering managers absolute night
terrors lately.

6
00:00:17,646 --> 00:00:18,826
Ah, let me guess, mate.

7
00:00:19,446 --> 00:00:24,306
Someone left a subagent looping in
extended thinking mode overnight and burned

8
00:00:24,346 --> 00:00:27,706
through the entire monthly API budget
before breakfast?

9
00:00:28,431 --> 00:00:29,251
Literally that!

10
00:00:29,811 --> 00:00:34,351
You have autonomous subagents or custom
prompts kicking off heavy reasoning tasks,

11
00:00:34,732 --> 00:00:37,651
and suddenly your token costs go straight
into orbit.

12
00:00:37,848 --> 00:00:42,247
It is a proper headache, especially across
enterprise teams where everyone is

13
00:00:42,287 --> 00:00:44,208
writing their own custom workflows.

14
00:00:45,127 --> 00:00:49,428
But, uh, Claude Code version two point one
point two hundred and sixty seven

15
00:00:49,668 --> 00:00:52,128
actually just dropped a fix for this exact
mess.

16
00:00:52,327 --> 00:00:54,347
The max effort level setting, right?

17
00:00:54,474 --> 00:00:54,954
Spot on.

18
00:00:55,595 --> 00:01:00,335
You can set maxEffortLevel at the top
level or per model under modelSettings.

19
00:01:01,054 --> 00:01:05,774
And what it does is enforce a strict,
unbreakable upper ceiling on extended thinking

20
00:01:05,835 --> 00:01:06,175
effort.

21
00:01:06,527 --> 00:01:07,007
Wow.

22
00:01:07,127 --> 00:01:09,607
And that applies across all cloud
providers, doesn't it?

23
00:01:09,668 --> 00:01:11,827
Like, not just direct Anthropic API?

24
00:01:12,054 --> 00:01:13,355
Yeah, every single one.

25
00:01:13,834 --> 00:01:15,855
Bedrock, Vertex, and Google Foundry.

26
00:01:16,654 --> 00:01:20,894
It caps the reasoning effort level on
every provider, but, and this is the crucial

27
00:01:20,974 --> 00:01:25,314
bit, users can still pick a lower level if
they're just running a lightweight task.

28
00:01:25,554 --> 00:01:30,154
So if a team lead sets the global ceiling
to, say, medium, an individual developer

29
00:01:30,214 --> 00:01:34,314
or an automated prompt can't pull a fast
one and request maximum high effort

30
00:01:34,354 --> 00:01:34,715
reasoning?

31
00:01:35,069 --> 00:01:36,009
Exactly, mate.

32
00:01:36,549 --> 00:01:38,529
It just clamps it right down to the cap.

33
00:01:38,909 --> 00:01:43,050
No more surprise five hundred dollar bills
because a subagent decided to contemplate

34
00:01:43,090 --> 00:01:45,989
the meaning of life on a basic code
refactor.

35
00:01:46,722 --> 00:01:48,282
That's huge for governance!

36
00:01:49,101 --> 00:01:51,942
But, wait, how does this interact with
custom skills?

37
00:01:52,462 --> 00:01:56,021
Because I know a lot of devs put effort
colon declarations right in the YAML

38
00:01:56,081 --> 00:01:58,002
frontmatter of their subagents or
commands.

39
00:01:58,161 --> 00:02:02,341
Ah, yeah, so previously, if a model's
default effort level was pinned,

40
00:02:02,781 --> 00:02:06,281
that frontmatter effort flag could end up
getting completely ignored.

41
00:02:07,002 --> 00:02:10,882
This update fixes that bug too, so your
frontmatter effort settings work as

42
00:02:10,922 --> 00:02:14,901
expected, provided they don't breach that
maxEffortLevel ceiling you set.

43
00:02:15,179 --> 00:02:19,839
Okay, so it respects the requested effort,
unless it exceeds the guardrail?

44
00:02:19,994 --> 00:02:20,495
You got it.

45
00:02:20,934 --> 00:02:24,795
It keeps the guardrail solid while letting
your subagents operate normally

46
00:02:24,875 --> 00:02:25,414
underneath it.

47
00:02:25,577 --> 00:02:26,717
That is so clean.

48
00:02:27,237 --> 00:02:31,097
And, er, while we are under the hood,
there was another CLI flag in this release

49
00:02:31,137 --> 00:02:32,837
that caught my eye for prompt engineering.

50
00:02:33,317 --> 00:02:35,937
The system prompt snapshot off flag?

51
00:02:36,280 --> 00:02:41,299
Oh, mate, that one is a total gem if you
are actively tweaking prompts.

52
00:02:41,859 --> 00:02:46,960
Normally, Claude Code records and reuses
the conversation's snapshot prompt to save

53
00:02:47,000 --> 00:02:48,920
on tokens via prompt caching.

54
00:02:49,344 --> 00:02:52,504
Right, so if you're iterating on system
prompt text mid session,

55
00:02:52,964 --> 00:02:55,644
it stays locked to whatever prompt started
the conversation?

56
00:02:55,786 --> 00:02:56,326
Exactly.

57
00:02:57,167 --> 00:03:02,366
But when you pass dash dash system prompt
snapshot off, it bypasses that cached

58
00:03:02,386 --> 00:03:06,746
snapshot and renders the system prompt
fresh on every single request.

59
00:03:06,971 --> 00:03:12,712
So you can tweak your prompt text live,
run a command, see the immediate result,

60
00:03:12,772 --> 00:03:16,191
without having to kill the session and
start a brand new context every time?

61
00:03:16,429 --> 00:03:18,769
Fair dinkum, it saves so much time!

62
00:03:19,149 --> 00:03:23,089
You get instant feedback on your prompt
changes, and then once you're happy with the

63
00:03:23,149 --> 00:03:26,649
prompt, you just drop the flag and go back
to standard prompt caching.

64
00:03:26,911 --> 00:03:31,311
Man, between hard capping reasoning costs
with maxEffortLevel and having fresh

65
00:03:31,411 --> 00:03:36,071
prompt evaluation on the fly, this update
really feels like Claude Code growing up

66
00:03:36,132 --> 00:03:37,511
for production enterprise teams.

67
00:03:38,602 --> 00:03:39,542
Couldn't agree more, mate.

68
00:03:40,063 --> 00:03:43,602
It gives managers peace of mind on the
billing side while giving devs the

69
00:03:43,642 --> 00:03:45,122
flexibility they actually need.

70
00:03:46,083 --> 00:03:47,722
Alright, that's the wrap on this one.

71
00:03:48,122 --> 00:03:49,022
Catch you all next time!

72
00:03:49,083 --> 00:03:50,043
See ya!