1 00:00:00,000 --> 00:00:05,520 Brought to you by Jellypod AI, here is a quick question for anyone building with AI 2 00:00:05,578 --> 00:00:06,240 assistants. 3 00:00:06,400 --> 00:00:11,680 Have you ever noticed your terminal session slowing down to a absolute crawl, 4 00:00:11,800 --> 00:00:13,600 and you cannot figure out why? 5 00:00:13,863 --> 00:00:15,102 Oh man, all the time. 6 00:00:15,702 --> 00:00:18,242 You think you are just having a quick chat with your code base, 7 00:00:18,602 --> 00:00:22,022 and suddenly every single turn takes ten seconds to respond. 8 00:00:22,083 --> 00:00:25,123 mate, I, I, I literally had this happen last week. 9 00:00:25,523 --> 00:00:29,763 I was working on a project with eight custom workspace skills loaded up. 10 00:00:29,853 --> 00:00:31,283 Nothing crazy, right? 11 00:00:31,395 --> 00:00:32,963 Just some helper scripts. 12 00:00:33,123 --> 00:00:35,203 But I noticed this weird lag. 13 00:00:35,403 --> 00:00:40,803 Turns out, those eight custom skills were silently burning fifteen thousand context 14 00:00:40,872 --> 00:00:44,483 tokens on every single turn before I even typed a word! 15 00:00:44,699 --> 00:00:46,179 Fifteen thousand tokens? 16 00:00:46,899 --> 00:00:49,900 Just sitting there in the background prompt before you even ask it to do anything? 17 00:00:50,077 --> 00:00:50,477 Every. 18 00:00:50,877 --> 00:00:51,318 Single. 19 00:00:51,637 --> 00:00:52,037 Turn. 20 00:00:52,698 --> 00:00:55,637 It was like driving a ute with the handbrake pulled halfway up. 21 00:00:56,317 --> 00:01:00,477 You are paying for all that context overhead, and the model is getting bogged down 22 00:01:00,557 --> 00:01:04,097 reading system prompts for tools you might not even touch during that session. 23 00:01:04,375 --> 00:01:06,615 That context inflation is huge. 24 00:01:06,705 --> 00:01:10,775 Because people keep adding marketplace extensions and custom skills, 25 00:01:10,811 --> 00:01:15,175 thinking they only matter when you call them, but the model has to know they exist! 26 00:01:15,167 --> 00:01:15,967 Exactly right. 27 00:01:16,067 --> 00:01:21,847 But look, Anthropic just released Claude Code version 2 point 1 point 261, 28 00:01:21,937 --> 00:01:26,527 and according to their official changelog, they added a new command called skill 29 00:01:26,596 --> 00:01:32,767 doctor, specifically, and I quote, to show which loaded skills go unused and what 30 00:01:32,815 --> 00:01:35,407 they cost in context, so you can prune them. 31 00:01:35,536 --> 00:01:39,396 Wait, so it actually gives you a breakdown of the exact token count per skill? 32 00:01:39,417 --> 00:01:39,977 Yeah! 33 00:01:40,017 --> 00:01:44,217 You literally just type slash skill doctor right in your terminal session. 34 00:01:44,404 --> 00:01:49,017 It inventories every active skill, flags the ones whose definitions have not been 35 00:01:49,067 --> 00:01:53,337 invoked in your current conversation history, and prints out the exact context 36 00:01:53,377 --> 00:01:54,697 footprint for each one. 37 00:01:54,708 --> 00:01:55,988 That is so useful. 38 00:01:56,148 --> 00:02:01,028 So what is the actual workflow then once you run slash skill doctor? 39 00:02:01,148 --> 00:02:02,948 How do you prune things down? 40 00:02:02,958 --> 00:02:06,638 So, step one, you run the command and spot the bloat. 41 00:02:06,878 --> 00:02:11,918 In my case, I had a couple of massive SKILL dot md files that were basically full 42 00:02:11,975 --> 00:02:13,278 documentation dumps. 43 00:02:13,598 --> 00:02:17,998 So step two, if you do not need the skill for that project, 44 00:02:18,018 --> 00:02:22,318 you just disable or remove the bloated SKILL dot md definition. 45 00:02:22,531 --> 00:02:26,878 Or step three, which is my favorite, you tighten up the trigger descriptions in the 46 00:02:26,931 --> 00:02:27,598 frontmatter. 47 00:02:27,583 --> 00:02:31,983 Ah, right, so instead of shoving the entire instruction set into the system prompt 48 00:02:32,033 --> 00:02:36,383 upfront, you keep the description lean so Claude knows when to pull in the heavy 49 00:02:36,450 --> 00:02:36,863 stuff! 50 00:02:36,875 --> 00:02:37,435 Spot on! 51 00:02:37,542 --> 00:02:42,155 Keeps your prompt lean, saves your token budget, and speed snaps right back. 52 00:02:42,167 --> 00:02:44,807 That is a massive win for token optimization. 53 00:02:44,967 --> 00:02:50,247 And speaking of this 2 point 1 point 261 release, there are a few other killer 54 00:02:50,297 --> 00:02:52,247 quality of life upgrades in here too, right? 55 00:02:52,469 --> 00:02:53,309 Oh, absolutely. 56 00:02:53,829 --> 00:02:57,369 They added two new settings in dot claude slash settings dot json called 57 00:02:57,789 --> 00:03:01,590 bashOutputMaxChars and taskOutputMaxChars. 58 00:03:02,449 --> 00:03:06,109 You can bump those all the way up to 128,000 characters. 59 00:03:06,346 --> 00:03:07,426 Oh, thank goodness! 60 00:03:07,866 --> 00:03:10,726 No more truncated test logs right in the middle of a build error! 61 00:03:11,412 --> 00:03:11,631 Right? 62 00:03:12,152 --> 00:03:15,311 It used to dump the tail end to a text file if it overflowed, 63 00:03:15,871 --> 00:03:20,651 but now you can keep up to 128K inline before it offloads to disk. 64 00:03:21,212 --> 00:03:23,491 Huge when you are running big test suites. 65 00:03:23,667 --> 00:03:25,107 What about subagents? 66 00:03:25,147 --> 00:03:27,907 I saw something in the update notes about subagent prompts? 67 00:03:27,917 --> 00:03:31,437 Yeah, they added a flag, append subagent system prompt file. 68 00:03:31,624 --> 00:03:35,997 So if you have a massive prompt for a subagent that is too long to comfortably pass 69 00:03:36,024 --> 00:03:39,117 on the command line, you just load it straight from a file. 70 00:03:39,125 --> 00:03:39,605 Nice. 71 00:03:39,645 --> 00:03:43,765 And didn't they harden some of the safety prompts around bash commands as well? 72 00:03:43,750 --> 00:03:44,310 They did! 73 00:03:44,422 --> 00:03:47,190 They upgraded the dangerous rm safety prompts. 74 00:03:47,330 --> 00:03:52,710 Now it catches stuff like rm rf hidden inside positional parameters or nested inside 75 00:03:52,744 --> 00:03:54,630 double quoted sh c calls. 76 00:03:54,817 --> 00:03:59,430 So if a tool try to execute a dangerous deletion deep inside a subshell script, 77 00:03:59,476 --> 00:04:03,670 Claude Code flags it and asks for confirmation before blowing away your files. 78 00:04:03,786 --> 00:04:05,826 That is really smart. 79 00:04:06,726 --> 00:04:12,047 Between slash skill doctor saving context tokens, larger output buffers, 80 00:04:12,146 --> 00:04:18,347 and tighter safety checks, this 2 point 1 point 261 update is a really solid tune 81 00:04:18,406 --> 00:04:18,526 up. 82 00:04:18,625 --> 00:04:19,345 Too right, mate. 83 00:04:19,425 --> 00:04:25,025 Definitely run slash skill doctor on your workspace today and see what silent token 84 00:04:25,085 --> 00:04:27,025 eaters you have lurking in your prompt!