1 00:00:00,100 --> 00:00:04,340 This episode's brought to you by Jellypod AI, and we're diving straight in. 2 00:00:04,900 --> 00:00:09,020 Ever had an AI agent tell you "all tests pass" and then you find out it never 3 00:00:09,100 --> 00:00:09,780 actually ran them? 4 00:00:10,523 --> 00:00:11,083 Constantly. 5 00:00:11,583 --> 00:00:15,703 And that's the itch Claude Code 2.1.286 scratches. 6 00:00:16,243 --> 00:00:18,923 It dropped on September 30, 2026. 7 00:00:19,120 --> 00:00:19,660 Go on then. 8 00:00:20,958 --> 00:00:25,518 So, according to the changelog, when your project or user skills include one named 9 00:00:25,609 --> 00:00:29,118 verify, Claude is now told to run it right before committing. 10 00:00:29,209 --> 00:00:32,078 Except for docs only and tests only commits. 11 00:00:32,183 --> 00:00:32,463 Hang on. 12 00:00:32,903 --> 00:00:33,403 That's it? 13 00:00:34,263 --> 00:00:35,943 A skill with a particular name? 14 00:00:36,263 --> 00:00:37,243 That's the whole feature? 15 00:00:37,375 --> 00:00:38,255 That's the whole feature. 16 00:00:38,395 --> 00:00:40,175 And I think that's the clever part. 17 00:00:40,328 --> 00:00:40,488 Huh. 18 00:00:41,008 --> 00:00:45,368 Okay, but why's that better than me just writing "always test before you commit" in 19 00:00:45,428 --> 00:00:46,268 CLAUDE.md? 20 00:00:47,068 --> 00:00:47,568 I've done that. 21 00:00:48,348 --> 00:00:51,148 Mate, I've done that, um, a lot. 22 00:00:51,833 --> 00:00:53,753 Sure, and it works until it doesn't. 23 00:00:53,865 --> 00:00:58,393 Long session, lots of momentum, the context gets compacted, 24 00:00:58,433 --> 00:01:02,633 and that one line is, uh, it's just one sentence among thousands. 25 00:01:02,873 --> 00:01:07,114 I can't promise that's exactly what happens every time, but that's the failure mode 26 00:01:07,148 --> 00:01:08,274 people worry about. 27 00:01:08,374 --> 00:01:13,194 A named skill is a fixed hook the tool itself knows about, so it doesn't depend on 28 00:01:13,214 --> 00:01:14,794 the model remembering your note. 29 00:01:14,833 --> 00:01:15,233 Right. 30 00:01:15,260 --> 00:01:17,873 So what about a plain old git pre commit hook? 31 00:01:17,953 --> 00:01:19,313 That's been around forever. 32 00:01:19,473 --> 00:01:21,553 Hooks are great, but they're blind. 33 00:01:22,113 --> 00:01:24,693 The commit fails, and the agent sees an error after the fact. 34 00:01:25,453 --> 00:01:29,193 With verify, the way I read it, the check runs inside the loop. 35 00:01:29,773 --> 00:01:34,393 Claude sees the failure output in context, fixes the typo or the syntax error, 36 00:01:34,833 --> 00:01:37,393 runs it again, and only commits once it's clean. 37 00:01:37,417 --> 00:01:41,817 Ah, so it's less a bouncer at the door and more a mate checking your work before you 38 00:01:41,857 --> 00:01:42,617 leave the house. 39 00:01:43,825 --> 00:01:45,065 Yeah, that's better than my version. 40 00:01:45,208 --> 00:01:47,528 Okay, so what would setup look like? 41 00:01:47,576 --> 00:01:48,328 Walk me through it. 42 00:01:48,333 --> 00:01:53,533 You'd create a skill folder at .claude/skills/verify in your project. 43 00:01:53,720 --> 00:02:00,333 Or put it in ~/.claude/skills/verify if you want it across every project. 44 00:02:00,413 --> 00:02:02,413 Inside, you describe the pipeline. 45 00:02:02,493 --> 00:02:07,373 For a JavaScript project, that might be npm run typecheck, then lint, 46 00:02:07,405 --> 00:02:10,813 then npm test run in band so the tests don't fight each other. 47 00:02:10,913 --> 00:02:14,733 For Rust, cargo test quietly, then cargo clippy. 48 00:02:14,750 --> 00:02:16,190 And then you just, uh, test it. 49 00:02:16,222 --> 00:02:20,750 Something like "fix the auth route and commit." 50 00:02:20,750 --> 00:02:22,270 Exactly that kind of prompt. 51 00:02:22,390 --> 00:02:25,470 You should see Claude run verify before the commit. 52 00:02:25,550 --> 00:02:29,310 And if you ask it to touch only markdown, it should skip the run, 53 00:02:29,340 --> 00:02:30,990 because of that docs only exemption. 54 00:02:31,000 --> 00:02:31,800 Which makes sense. 55 00:02:31,903 --> 00:02:34,840 Nobody needs a type check for a typo in a README. 56 00:02:34,875 --> 00:02:35,195 Right. 57 00:02:35,368 --> 00:02:36,208 Now, I've got a worry. 58 00:02:37,688 --> 00:02:42,008 I once pushed a bad midnight update that nearly tanked a client's site, 59 00:02:42,128 --> 00:02:45,108 so I'm, I'm, I'm a bit twitchy about slow checks. 60 00:02:46,068 --> 00:02:48,008 What if verify takes forever? 61 00:02:48,350 --> 00:02:49,130 Then it's a problem. 62 00:02:49,930 --> 00:02:54,570 My take, and this is my opinion, not the changelog's: keep verify fast. 63 00:02:55,250 --> 00:02:58,950 If you've got an integration suite that takes ten minutes, don't put it in there. 64 00:02:59,610 --> 00:03:02,430 A turn can time out, and you'll be sitting there watching a spinner. 65 00:03:03,230 --> 00:03:05,510 Typecheck, lint, the quick tests. 66 00:03:06,170 --> 00:03:07,550 Leave the heavy stuff for CI. 67 00:03:07,625 --> 00:03:08,345 Fair. 68 00:03:08,409 --> 00:03:09,865 Fast and boring wins. 69 00:03:09,875 --> 00:03:10,475 Always. 70 00:03:10,515 --> 00:03:12,675 Oh, and the point release has a bunch of polish too. 71 00:03:12,733 --> 00:03:17,555 Permission prompts now show a count like "2 of 5" when several requests stack up. 72 00:03:17,583 --> 00:03:18,623 Oh, thank goodness. 73 00:03:18,903 --> 00:03:21,023 I hate not knowing how deep the pile is. 74 00:03:21,042 --> 00:03:22,242 There's also a safety one. 75 00:03:22,429 --> 00:03:27,922 If you're viewing a background agent's or teammate's transcript and type compact, 76 00:03:27,989 --> 00:03:32,802 clear, or rewind, it used to silently act on the main conversation. 77 00:03:32,942 --> 00:03:35,922 Now a dialog names the target and asks first. 78 00:03:36,618 --> 00:03:39,298 So you can't accidentally wipe the wrong session. 79 00:03:39,578 --> 00:03:42,858 That's, wow, that's a nasty one to have fixed. 80 00:03:42,958 --> 00:03:47,118 And in a subagent's view, control enter now moves its running command to the 81 00:03:47,176 --> 00:03:50,158 background, so your message gets read right away. 82 00:03:50,334 --> 00:03:54,718 Plus, if the API refuses the model your default resolves to, 83 00:03:54,764 --> 00:03:58,558 Claude Code retries once on the previous model of the same tier, 84 00:03:58,598 --> 00:04:00,318 instead of failing every single turn. 85 00:04:00,333 --> 00:04:06,413 So the big idea is just, name a skill verify, and the agent checks its own homework 86 00:04:06,493 --> 00:04:07,853 before it hands it in. 87 00:04:08,035 --> 00:04:08,435 That's it. 88 00:04:09,035 --> 00:04:13,995 One folder, one name, and the commit gets a gatekeeper that's actually in the room.