00:00:00So just this week, the Claw team have released Routines, a massive upgrade to the scheduling
00:00:05feature that runs your prompt on Anthropics Cloud infrastructure with custom connectors,
00:00:10environments and triggers them on a schedule via a post request or a GitHub event like creating
00:00:16a PR.
00:00:17But does this really mean the end for AI workflow tools like N8N and how expensive will this
00:00:23actually be?
00:00:24Hit subscribe and let's get into it.
00:00:29It looks like Anthropic are releasing more and more features that make use of their cloud
00:00:33infrastructure, from Managed Agents to UltraPlan.
00:00:37And now this Routines feature that adds three ways to run tasks in clawed code, loops to
00:00:43run in session tasks, schedules to run tasks at regular intervals on a specific date or
00:00:49time locally or remotely, and now Routines, which includes schedules, so this introduced
00:00:55remote schedules, but also add GitHub event triggers and API post calls.
00:01:01Let's go through how it works by running through two examples.
00:01:04The first task is to create a simple automation that gets data from three newsletters and gives
00:01:09me the best articles in Slack every day at 9am, and the second one is to automatically
00:01:16review pull requests as soon as they're created, adding inline comments with suggestions for
00:01:21improvements.
00:01:22Let's see how we can do these with Routines.
00:01:24So I'm going to do the first routine in the terminal via clawed code, and we can do that
00:01:28using the /schedule command.
00:01:31And from here we can continue through the steps to create our own routine.
00:01:34So if I hit create new, it's going to ask me a few more questions.
00:01:38Or what I'm going to do instead is, because I already know what I want, I'm just going
00:01:41to write the /schedule command, followed by a prompt for the routine I want to create.
00:01:46So create a daily 9am trigger that fetches RSS from JS Weekly, React Status, Node Weekly,
00:01:53and picks 10 good articles for YouTube videos to send me via Slack.
00:01:58And from here, Claude sets this up with a timezone, environment, with a draft prompt.
00:02:03So once that's done, it creates a remote trigger by default, which I can view in the
00:02:06browser or in the Claude app.
00:02:09Now I'm sure if I wanted to, I could create a local trigger, but for now, let's take
00:02:12a look at the remote routine.
00:02:13And here it is under the Routines option, I can click on it to see that it repeats every
00:02:18day at 9am, and has used my Slack connector with the instructions from Claude code.
00:02:24Now there are a few things that it's done that I'm not a fan of.
00:02:26First of all, it automatically connected a GitHub repo, which is unnecessary for this
00:02:31specific routine.
00:02:32So I'm going to change that by moving it from here.
00:02:34And also I'm going to edit this prompt to tell it not to use horizontal rule dividers since
00:02:39they cause invalid block validation errors inside Slack.
00:02:43It's important to note that before you start creating a routine inside Claude code, you
00:02:47need to have your connectors already, and also make sure that you write a prompt in such
00:02:51a way that it doesn't need to ask for permissions since this runs autonomously kind of in a hands-off
00:02:57mode.
00:02:58Luckily, Claude code has automatically written the prompt for me, so I don't have to review
00:03:02it.
00:03:03Now, before I continue with this prompt, it's important to note that I've used the newsletters
00:03:06environment instead of the default one.
00:03:09And the reason for this is because I've done this before and I know what went wrong.
00:03:12But in all seriousness, these URLs, so this RSS for JavaScript Weekly, React Status and
00:03:17Node Weekly are URLs that Claude wouldn't allow the bash tool to run curl on or to fetch
00:03:23from because it has a security layer that prevents all outbound network requests for bash.
00:03:29So in order to get around that, I created a new environment by clicking here, changed
00:03:33to custom and added the list of allowed domains.
00:03:37If you use the web fetch tool instead of bash for fetching web information, then that doesn't
00:03:41have that restriction because these calls are going through anthropics infrastructure.
00:03:45So I'm sure they have some extra security to make sure no prompt injection is happening.
00:03:49Okay, now that we're done with this, I'm going to close and do a test run of this routine.
00:03:54So now the workflow started, we can see over here and we can click here to see exactly
00:03:59what's going on.
00:04:00So here it clones the repo if there is one, but because there isn't, it's just set up the
00:04:03cloud container and it started to fetch the RSS feed.
00:04:06So it's done it for all three URLs using the bash curl tool and it seems like it's hit DNS
00:04:12issues.
00:04:13So it's fetching with the web fetch tool anyway, but it's good to know how to use this approach.
00:04:17It's getting the latest articles for the newsletters and it seems to have failed on Node Weekly,
00:04:21even though when I said to this earlier it worked before.
00:04:24Okay, so now it's found the top 10 articles that would be great for a YouTube video.
00:04:28And if we scroll down to the bottom, we can see it sent the message to me in the newsletters
00:04:32Slack channel.
00:04:33And if we go to Slack, we should see the message in the career channel with a link to all the
00:04:37articles, which is very nice.
00:04:39Now this brings us on nicely to how routines are priced.
00:04:42So right now routines are in research preview and can only be used with a Pro, Mac, Team
00:04:46or Enterprise subscription and draw from your subscription usage limits the same way an interactive
00:04:51session would.
00:04:52However, routines have an additional daily cap, which if we look at what it is for a Pro
00:04:57subscription, we can see over here it's five routines per 24 hours.
00:05:01So I think this is Anthropic's way of preventing you from running loads and loads of routines
00:05:05every day, or to kind of gauge how much people are using routines on their infrastructure
00:05:10to make sure too many resources are used.
00:05:13But this test routine run that I've made just now doesn't count towards my five daily
00:05:18limits.
00:05:19So I can prove that to you by running the exact same routine again.
00:05:23And we can click here to see that there have been two runs.
00:05:26So I'm going to close this sidebar here.
00:05:28I'll let this go for a bit.
00:05:30And now that it's going, I'm going to ask it to stop.
00:05:32And once it stopped, we can see that two routines have run.
00:05:35So they've been triggered manually, but I still have only one routine used in my daily
00:05:39limit.
00:05:40Now I'm going to create the next routine inside the cloud desktop app.
00:05:44And I'll explain why in a few moments, but you can already see that it's over here.
00:05:47I created it earlier, but let me show you how I did it.
00:05:50So first I clicked on routines over here, which you could probably guess, and then I clicked
00:05:54new routine, a remote one.
00:05:55I gave it a name and a prompt.
00:05:57And because this routine is to automatically review a pull request as soon as it's created,
00:06:01I selected a repo for it to view pull requests from, in this case, my dot files one.
00:06:06From here, we can see we have three trigger options, schedule, GitHub event and API.
00:06:10From cloud code CLI, you can only do schedule.
00:06:13So if you want to run GitHub event or API, you'd have to use the desktop app.
00:06:18And that's why we're using it here.
00:06:19So I'll click GitHub event and when API is opened.
00:06:22So we'll hit add trigger and we'll close the existing connectors.
00:06:25Now this is pretty much how I created that routine, but with one major tweak, I added
00:06:31my own custom skill, which I got from this repo.
00:06:35Now because Claude's code in the cloud creates a new instance of Claude code, it doesn't
00:06:40have access to your local skills or settings or hooks.
00:06:44So if you want to give it access to that, you need to do so through a repo and I've done
00:06:48it through this routine and for repository.
00:06:51So inside this, I have a dot Claude folder, which all it contains is a settings JSON file
00:06:56and my PR review skill, which I want to use in that routine.
00:07:00And all this does is it triggers a hook.
00:07:02So if you go back to Claude and hit on settings JSON, all that does is triggers a hook to make
00:07:07sure that this repo exists inside the Claude code cloud instance.
00:07:12So that's the routine and repo and copy all the skills from that repo into the Claude skills
00:07:17directory inside the cloud version of Claude.
00:07:20So if you close this and open my actual routine, then take a look at the prompt.
00:07:24You can see I've got some extra guard rails in place to make sure that this actually happened.
00:07:29Now I've already run this to test it and I had some issues.
00:07:32So let's try it again and hopefully it will work first time.
00:07:35So inside my dot files repo, I'm going to create a new pull request and a few seconds later,
00:07:40the PR review web hook starts running.
00:07:42I'm going to hide this to make it bigger and we can see it's run the prompt.
00:07:46It's got the pull request and I want to see if it's actually closed the repo.
00:07:50So if you go down here, it's cloned two repos, it can see the PR review skill and now it's
00:07:54running the PR review skill on that PR.
00:07:58And I forgot to add my GitHub token, which is something I explicitly asked for in the skill.
00:08:03But it's used the GitHub MCP tool instead, which is great.
00:08:06It's noticed that diff is minimal and it finished reviewing the PR.
00:08:09So if we go back to the PR, we can see Claude has added a comment saying automated review
00:08:13complete and no issues found.
00:08:16But unfortunately, because this wasn't a test run, this is actually used up one of my five
00:08:20runs a day.
00:08:21And so I could only run five automated PR review checks every single day, but it looks like
00:08:26you get 15 runs in the 20x max plan.
00:08:29So will I be using more of Claude routines in the future?
00:08:32Very unlikely.
00:08:33I mean, for what it is, it's very expensive.
00:08:36Okay.
00:08:37If I had a lot of money and I was paying for multiple 20x max accounts, then sure I'd run
00:08:43routines all the time.
00:08:44But right now it's much cheaper for me to configure webhooks with something like Hermes
00:08:48agent, or to use Multica with Hermes agents on a cheaper model like GLM 5.1 or a GPT coding
00:08:55model.
00:08:56However, I will admit it's so much easier to create multiple routines using the setup I
00:09:01showed earlier than to do everything from scratch using Hermes and Multica.
00:09:05I mean, that would take a long time, even with an LLM to help me out.
00:09:09But even if I did have a lot of money, I wouldn't say routines are for every type of automation.
00:09:14As you saw earlier, you just give a routine one prompt and it's supposed to do everything
00:09:19from that.
00:09:20I mean, okay, yes, if you are very technical, you could add extra skills and extra prompts
00:09:25to a repo, get the routine to clone that repo, and then chain these prompts together to have
00:09:30a more complicated routine.
00:09:32But it still costs a lot of money.
00:09:34You can't get past that.
00:09:36I will admit, however, that it feels like Anthropic are leading onto something bigger with all
00:09:41these cloud features that they're adding recently.
00:09:43And I'm really looking forward to seeing what it is.
00:09:45Speaking of cloud features, if you want to see how quick and easy it is to create and
00:09:50deploy an agent in minutes using the cloud managed agents feature, then check out this
00:09:55video.