1
00:00:00,000 --> 00:00:01,279
Welcome to the show, tech heads!

2
00:00:01,439 --> 00:00:05,919
I'm Lachlan Reed, and today we are looking
at a massive shift in how we build with

3
00:00:06,039 --> 00:00:06,959
LLMs.

4
00:00:07,066 --> 00:00:11,999
James, have you seen the Claude Code
v2.1.157 release?

5
00:00:12,111 --> 00:00:16,644
They are completely bypassing marketplace
registries by letting you load custom

6
00:00:16,700 --> 00:00:19,684
execution scripts directly out of a local
directory.

7
00:00:19,964 --> 00:00:23,444
I am talking about the
dot-claude-slash-skills folder.

8
00:00:24,000 --> 00:00:24,960
I saw that!

9
00:00:25,428 --> 00:00:26,322
It's brilliant.

10
00:00:26,402 --> 00:00:30,801
Instead of jumping through enterprise API
gateway hoops or waiting for a registry

11
00:00:30,854 --> 00:00:36,406
approval, Claude just scans your local
dot-claude-slash-skills directory at startup.

12
00:00:36,619 --> 00:00:41,447
If there's an executable file or script in
there, Claude registers it as a tool.

13
00:00:41,607 --> 00:00:47,284
You can bootstrap a new one instantly
using "claude plugin init" and then pass the

14
00:00:47,332 --> 00:00:48,325
name of your plugin.

15
00:00:49,000 --> 00:00:51,717
It is a game changer for local developer
velocity.

16
00:00:52,775 --> 00:00:56,921
Back in my day, we had to rebuild the
whole container to test a tool.

17
00:00:57,121 --> 00:01:02,601
Now, if you change your script's logic
mid-session, you just run slash-reload-skills

18
00:01:02,668 --> 00:01:07,240
right inside the Claude terminal, and boom
-- it hot-swaps the definition on the

19
00:01:07,340 --> 00:01:07,720
fly.

20
00:01:07,960 --> 00:01:09,637
No restarts required.

21
00:01:10,000 --> 00:01:13,917
Slash-reload-skills is massive for tight
feedback loops.

22
00:01:14,117 --> 00:01:17,354
But Lachlan, think about the git
implications here.

23
00:01:17,487 --> 00:01:22,393
If teams start committing these custom
dot-claude-slash-skills scripts directly to

24
00:01:22,439 --> 00:01:27,590
source control so everyone on the team has
the same custom tools...

25
00:01:27,797 --> 00:01:30,067
doesn't that open up a pretty nasty attack
vector?

26
00:01:31,000 --> 00:01:31,959
Oh, absolutely.

27
00:01:32,743 --> 00:01:35,881
It is a massive security blindspot if you
are not careful.

28
00:01:35,988 --> 00:01:40,927
Think about it: you git pull a public
repo, it has a malicious script sitting inside

29
00:01:41,007 --> 00:01:45,096
dot-claude-slash-skills, and the moment
you spin up Claude,

30
00:01:45,149 --> 00:01:47,975
it executes that script with your local
shell privileges.

31
00:01:48,135 --> 00:01:51,255
It's basically arbitrary code execution
waiting to happen.

32
00:01:52,000 --> 00:01:52,717
Exactly.

33
00:01:52,744 --> 00:01:56,958
We're essentially treating our LLM tool
directories like npm packages,

34
00:01:57,018 --> 00:01:59,429
but without the registry-level security
scanning.

35
00:01:59,589 --> 00:02:04,466
If you're committing executable plugins to
source control, you absolutely have to

36
00:02:04,506 --> 00:02:09,658
treat that folder with the same paranoia
as a pre-commit hook or a CI/CD pipeline

37
00:02:09,715 --> 00:02:10,217
script.

38
00:02:11,000 --> 00:02:11,959
Spot on, mate.

39
00:02:12,092 --> 00:02:17,003
It's a classic tradeoff -- ultimate local
flexibility, but you've got to watch your

40
00:02:17,083 --> 00:02:17,763
step.

41
00:02:17,835 --> 00:02:20,966
That's our quick take on Claude Code's
local skills.

42
00:02:21,039 --> 00:02:22,445
Catch you in the next one!