06:13Conor Martin · Tutorial
Claude Design Can Make Marketing Videos Now
A six-minute walkthrough of using Claude Design and HyperFrames to turn a design system into a rendered AI marketing video.
April 27thThe argument in one line.
n8n workflows called from Claude Cowork give you structured, auditable, token-efficient automation without the opacity, compliance gaps, or execution ceiling that come from chaining skills alone.
Read if. Skip if.
READ IF YOU ARE…
- You already have an n8n subscription and Claude Cowork but have not connected them yet.
- You build AI automations at a company where compliance, audit logs, or regulated data handling is a real constraint.
- You want to extend what Cowork can reach — YouTube, scraping, external APIs — without burning through your Claude token limit.
- You want to build n8n workflows faster by letting Claude draft the scaffold, then cleaning up manually.
SKIP IF…
- You have no n8n subscription — the token-saving argument assumes you already have n8n executions to spend.
- You are looking for n8n fundamentals — this assumes you already understand nodes, webhooks, and credentials.
The full version, fast.
Most people treat Claude Cowork and n8n as separate tools. This episode argues they are better together: Cowork handles the conversational front end, n8n handles structured execution. Connecting them takes 60 seconds via the built-in connector. Any n8n workflow then becomes callable from a chat window, trading token-hungry skill chains for the execution quota you already pay for. The email classifier demo shows Claude drafting, revising, and validating a 14-node workflow iteratively. The sharpest insight is compliance: Cowork skill histories are not auditable, but n8n execution logs are — which matters the moment a regulator asks what the AI was doing.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Who's talking.
00:00hostn8n podcast host
02:00guestRyan Nolan
Where the time goes.
00:00 – 02:23
01 · Introduction and episode overview
Host sets up Ryan Nolan's background and scopes the episode: Cowork + n8n connection, email classifier build, token savings, enterprise compliance, hackathon story.
02:23 – 04:21
02 · Connecting n8n to Claude Cowork
60-second connector setup: Customize > Connectors > browse > n8n > Allow. Permission granularity settings per action type.
04:21 – 10:27
03 · Calling APIs from Cowork via n8n
Webhook > HTTP Request > Respond to Webhook demo. Claude calls a Rapid API YouTube transcript fetcher, grabs just the video ID, returns clean transcript. Execution log shows 2-second run time.
10:27 – 14:51
04 · Building an email classifier workflow
Ryan asks Claude to build an n8n email classifier from scratch. Claude looks up SDK reference and nodes simultaneously. Discussion: use Opus not Sonnet for workflow creation.
14:51 – 20:10
05 · Tips for prompting and iterating
How Ryan prompts Claude for workflow builds: throw in a lot of context, say ask me if unsure, then clean up the mess. Cowork as a multiplier — better the more you know n8n.
13:42 – 17:10
06 · Compliance and audit logging
In regulated industries you must log every AI interaction. Skills don't produce structured logs; n8n execution history does — inputs, outputs, pinned data, node-by-node.
24:10 – 27:10
07 · Saving tokens with n8n workflows
Chaining skills burns tokens. A single skill that calls an n8n workflow uses n8n's execution quota instead. Token-saving pattern: replace multi-skill chains with one n8n webhook call.
32:31 – 34:08
08 · Enterprise security and skills vs workflows
Skills live per-machine — a compliance problem. n8n workflows are centralized and auditable. Non-technical domain experts can use Cowork chat front end without understanding n8n internals.
36:48 – 39:54
09 · Reviewing the built workflow
Ryan loads the Claude-generated 14-node classifier in n8n: text classifier, error handling on 5 nodes, Gmail send-and-await with 24-hour timeout, spam branch, was-approved branch. OpenAI credentials used instead of Claude — a common mistake.
39:41 – 42:10
10 · Quick Google credential setup
New Sign in with Google flow: one click, choose account, allow, done in 5 seconds. Replaces the old Google Cloud Platform documentation slog.
42:10 – 44:01
11 · Execution history and debugging
Copy to editor brings failed execution into workflow with data pinned. Use this to debug hallucinations — see exactly what went into the AI node and what came out.
44:01 – 46:45
12 · The hackathon story
A team with zero n8n experience built a driver's license verification workflow in two days: webhook > visual model checks > OCR > name match > approve/reject. Now in production.
46:45 – 51:43
13 · Advice for getting started
Observe before automating. Identify intersection of feasibility and value. Hold their hand through the first build. Test 50-200 runs. Expand department by department.
51:43 – 54:42
14 · Running with David Goggins
Ryan ran his first marathon in Las Vegas, started cramping at mile 19, saw Goggins run past. Goggins waited at the finish line to acknowledge everyone he paced — a week after running Moab 240.
54:42 – 55:30
15 · Where to find Ryan
Ryan and Nolan Data Science YouTube — 17-hour n8n course, 100+ videos. Free School group with weekly Wednesday calls.
Lines worth screenshotting.
- Connecting n8n to Claude Cowork via the built-in connector takes under 60 seconds — click the connector, choose Allow, done.
- YouTube is blocked inside Claude Cowork natively; any workflow that needs YouTube data must go through an n8n webhook calling an external API.
- Chaining Cowork skills multiplies token spend; replacing a chain with a single n8n workflow call uses your n8n execution quota instead of your Claude limit.
- Skills live on individual machines; n8n workflows are centralized — which becomes a compliance issue the moment someone asks who ran what.
- Cowork skill histories are not auditable logs; n8n execution history with pinned inputs and outputs per node can stand up to regulator scrutiny.
- Claude defaults to a less capable model when building workflows unless you explicitly switch to Opus — and the node quality difference is visible in the output.
- The always-allow permission for updating n8n workflows is the dangerous one — Claude will revise working workflows when trying to fix broken ones.
- n8n Sign in with Google turns a multi-step Google Cloud Platform credential setup into five seconds.
- When Claude builds an n8n workflow from scratch, expect at least one wrong node choice — but the scaffold still saves hours of manual construction.
- A team with zero n8n experience built a production driver's license verification workflow in two days — the visual node interface lowers the ramp for domain experts without coding backgrounds.
- Automators are aggressively lazy: they spend 30 minutes automating a 30-second task so it never has to happen again.
- The correct enterprise automation rollout sequence is observe first, identify high-value processes second, automate third — never automate everything at once.
n8n as the structured backbone Claude Cowork cannot be.
WHAT TO LEARN
Cowork is a conversational front end; n8n is the auditable, token-efficient execution layer behind it — and the combination is more useful than either tool alone.
02Connecting n8n to Claude Cowork
- Connecting n8n to Claude Cowork takes under 60 seconds through the built-in connector — most people who own both tools have never done it.
- Permission granularity matters: needs-approval, always-allow, and block are set per action type, not globally — tune them before your first real workflow.
03Calling APIs from Cowork via n8n
- YouTube and many external APIs are blocked natively inside Cowork; n8n workflows are the bridge, letting Claude call any external service through a webhook.
- Claude extracts just the video ID from a full YouTube URL before passing it to the API — you do not need to preprocess inputs manually.
04Building an email classifier workflow
- When Claude builds an n8n workflow from scratch, expect at least one wrong node choice — use a more capable model for generation, then verify every node manually before production.
- Asking Claude to flag questions and rate its own prompts before you review reduces hallucination in the generated workflow nodes.
05Tips for prompting and iterating
- Use a more capable model to generate and revise complex n8n workflows, then switch to a lighter model for simple tasks to control token spend.
- Throw as much context as possible into the initial workflow generation prompt — you will spend less time correcting than if you start sparse.
06Compliance and audit logging
- Cowork skill histories are not auditable logs — n8n execution history with pinned inputs and outputs per node is what compliance teams actually need.
- In regulated industries, every AI interaction must be logged with its input, output, and the human-in-the-loop decision — n8n makes this visual and exportable.
07Saving tokens with n8n workflows
- Chaining multiple Cowork skills multiplies your token spend on every run; replacing chains with a single n8n workflow call uses your n8n execution quota instead.
- If you already pay for an n8n subscription, you have an execution quota you are likely underusing — routing work through workflows instead of skills taps that quota.
08Enterprise security and skills vs workflows
- Skills live only on the machine that created them; centralized n8n workflows are accessible to the whole team and can be audited.
- Non-technical domain experts can use Cowork as a chat front end without understanding n8n internals — lowering the barrier to AI automation across a company.
09Reviewing the built workflow
- The always-allow permission for updating n8n workflows is the most dangerous one — Claude will revise working workflows when trying to fix broken ones.
- After any AI-generated workflow, check the model credentials first — Claude frequently defaults to OpenAI credentials instead of your configured Claude credentials.
10Quick Google credential setup
- n8n Sign in with Google eliminates the Google Cloud Platform documentation process entirely; what used to take many steps now takes five seconds.
11Execution history and debugging
- The copy-to-editor feature opens a failed run with all data pinned so you can replay and debug without re-triggering the webhook.
- When an AI node hallucinates in an n8n workflow, you can see exactly what data entered the node and what came out — impossible to do inside a Cowork skill.
12The hackathon story
- Domain experts with no coding background can learn to build useful n8n workflows in two days when the problem they are solving is one they understand intimately.
- Visual workflow tools lower the barrier for domain-expert-built automation — and the resulting workflows are more accurate because the builder knows the problem.
13Advice for getting started
- The correct sequence for rolling out AI automation is observe first, identify high-value low-complexity processes second, automate third.
- Test any automation workflow 50 to 200 runs before declaring it done; errors compound at scale, not one at a time.
- Universal error handling — a node that fires when anything fails — is the first thing to add and the first thing most demos skip.
Terms worth knowing.
- Cowork
- Anthropic's Claude desktop app with skills and connectors that lets Claude perform actions on your computer and connect to third-party services.
- Connector
- A built-in integration inside Claude Cowork that links Claude to a third-party service so Claude can call its APIs directly from chat.
- Human-in-the-loop node
- An n8n node that pauses a workflow and sends an approval request before proceeding — the workflow resumes only after a human approves or declines.
- Text Classifier node
- An n8n node designed for routing inputs into labeled categories; more structured than a basic LLM chain for classification tasks.
- Execution history
- n8n's per-run log of every node's input and output, used for debugging and compliance documentation.
- Skills
- Reusable task definitions inside Claude Cowork that let Claude perform specific actions — but they live on the individual machine that created them.
- Copy to editor
- An n8n feature that opens a failed execution in the workflow editor with all data pinned, letting you replay and debug without re-triggering the workflow.
Things they pointed at.
54:42channelRyan and Nolan Data Science YouTube
55:10linkRyan Nolan free School group
Lines you could clip.
07:13
“Trust but verify — that's the name of the game with these AI systems.”
23:13
“Errors in automations don't happen one at a time. They happen at scale.”
26:28
“Automators are aggressively lazy. You'd rather spend thirty minutes so you never have this problem ever again.”
33:19
“There are people at every company brilliant at their domain and struggling with technology. Cowork is a chat interface — they understand that.”
51:19
“If something works, you don't always have to change it.”
Where the conversation goes.
00:00 – 04:21steadySetup and connection
04:21 – 10:27denseAPI calling demo
10:27 – 39:54denseEmail classifier build
24:10 – 34:08denseToken savings and enterprise
36:48 – 44:01steadyWorkflow review and credentials
42:10 – 46:45denseCompliance, debugging, hackathon
46:45 – 51:43steadyGetting started advice
51:43 – 55:30sparsePersonal closing
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
00:00Click on this plus icon over here, then it's gonna say Claude wants access to your NAN instance. Choose allow. You see success.
00:06Open up Claude once again. Connect it to NAN. Literally, that took thirty seconds, if not that.
00:11In cohort, you can chain together skills, but this is gonna eat up your tokens a lot. If you already have an NAN subscription, build out a NAN workflow. Instead of having all these skills run, all your tokens, and hitting those cloud limits, why don't you use the executions that NAN gives you?
00:26Sign in with Google. Choose your account. Account connected.
00:29Five seconds. I remember my first time setting this up. I was like, man, there's so many different steps.
00:33This is super technical. Now there's, like, no reason for anyone to watch that YouTube video how to set up Gmail credentials. We're gonna be learning about how to use CloudCourc
00:41with NAN. We're gonna go through a real practical workflow, something super basic that anyone can start today, and we're gonna go over 10 to 11 different benefits of utilizing this. You've probably seen Claude Cowork all over your feed.
00:52Anthropics desktop app with skills and connectors that lets Claude actually do things on your computer. And the question that every NA and builder is asking is how do I use this in my existing workflow? Ryan Nolan runs the YouTube channel, Ryan and Nolan Data Science and Works at Bill, where he's been deploying AI automations across departments.
01:10He's the right person to answer this, and today's episode is the Playbook. He shows us how to connect Cowork and Claude to n eight n in less than sixty seconds, build an entire email classifier with the Gmail human in a loop node and error handling. The token saving patterns that most people miss, every n eight n workflow that you call from Cowork doesn't.
01:28The new sign in to Google credentials that turns the hardest part of n eight n setup, the Google Cloud documentation slog into five seconds. Plus the enterprise side. Why compliance teams care about the execution history, why skills living on an individual computer is a compliance problem, and the hackathon story where a team that had never touched an eight in built a production driver's license workflow in two days.
01:51Let's get into it. Ryan, excited for having you on the show, brother. So what are we gonna be learning today?
01:55We're gonna be learning about how to use Claude CoWork with NAN. We're gonna go through a real practical workflow, something super basic that anyone can start today, and we're gonna go over 10 to 11 different benefits of utilizing this. I'm really, really excited about it.
02:08Me too. Uh, this is something that is a very hot topic using Cowork and NN together so that you can be more productive. It's something that I'm interested in, and I'm sure a lot of people that are watching this are gonna be interested as well.
02:19So feel free to share your screen when you're ready, and let's get into it, shall we? Okay. So here I have our Claude desktop app.
02:25You need to have this installed to be able to use Claude CoWork correctly. And then I'm here on this center tab. Now there's a lot to the Claude Cowork app.
02:34Obviously, we cannot talk about it all today, but we're gonna take you step by step. The first thing we're gonna do is set up any n directly in over here. So the way to do that is you go over here to this customize icon.
02:44Now
02:45fully transparent, Claude changes their icons and locations all the time. So if you don't see this in the future, I promise you it's gonna be pretty easy. So jump into customize.
02:54Then what we're gonna do is look for our connectors. And in over here, we're gonna click this plus icon, and we're gonna browse different connectors.
03:03And connectors are just integrations into other third party applications, so that way you can start sending data that way or grabbing data for any sort of skills. So NAN has a built in, uh, connector, which makes it really easy to do.
03:17And all we're gonna do is go into the connectors over here and search NAN. Now you can see I don't have this activated right now, but I'll show you how to set this up in literally only sixty seconds. So click on this plus icon over here, then it's gonna say, Claude wants access to your n e n instance.
03:32Obviously, choose allow. You see success. You'll be redirected back into the client.
03:38Open up Claude once again, connected to NAN. Literally, that took, what, thirty seconds, if not that.
03:44Yep. And what's what's really great is you can actually determine if you want full access to NAN or just go through different settings. So So you can see, for example, getting execution data, it needs approval.
03:56Uh, you can go over here and say, always allow, or you can just keep this as new to approval, or you can say, hey. I don't want any of this execution data to show up within Claude Coark, you and can turn this off.
04:06Now, obviously, I'm not gonna go through every single one of these settings. But based off of your own security preferences, you can either have these as new needs approval, uh, automatic approval, or completely block it. So as you see, about a minute to set this up.
04:18Super easy. But you're like, okay. That's great.
04:21How do we actually work with some sort of a workflow here in CloudCore? Again, let's, uh, start with the basics.
04:28So I'm gonna go back over here, and we're gonna go straight in this chat panel. And I'm gonna just show you a very quick demo.
04:35One of the things I do for my YouTube channel is I try to get the transcripts from a specific video, and then I'll turn those into time stamps. Now I'm not gonna show that full workflow today. I'm just gonna show you the very basics of how do we call an external API with the help of any n.
04:51So in any n, I have this basic workflow, and I would share it, but, like, it literally takes you two seconds to make, or you can ask Claude Cowork itself to make it.
05:00We have a webhook that over here. Then we have our HTTP request and then respond to webhook. Literally, just three nodes to get started with this.
05:09Obviously, you can go as complicated as you want, but let's start with the very basic demo. Our webhook is our chat, API request, respond back over.
05:19So then I'm gonna go back over here. And how am I gonna actually call this workflow? So this is called Cowork YouTube transcript.
05:28And all I'm gonna say in over here is, can you run the transcripts?
05:37Can you run the coworker transcript workflow? And I'm just gonna paste in a YouTube URL.
05:43I have a collectibles channel, so I'm gonna actually paste in a video over here that I didn't do time stamps on. And here is this specific video. And if you worked with Clowork for a while, you'd know that YouTube is blocked directly in over here.
05:55So you're gonna have to use some sort of API to analyze a video or get data from the video. Obviously, I want the transcripts, so I'm just gonna run this directly like that. And for this, grabbing the actual transcripts, because it does get blocked, are you using, like, Appify or some other service to do that?
06:10Yeah. So I'm using Rapid API on this side of things. And what you'll notice over here as well, because of our permission settings, it says Claude wants to use the execute workflow from NaN.
06:20So it's probably blurred, but it has our workflow ID. It has the type webhook, which I just showed you, and then it also has the body. And all it is requesting is this specific video ID.
06:29So Claude is smart enough to take this YouTube URL and then just grab the specific ID that we particularly want. So I'm gonna click allow for this task, and it says Claude wants to use get execution from NAN.
06:41Again, you could always allow these. I'm just doing this for this demo purpose just to show you guys the first time that you're gonna get these directly in here. I'm gonna click always allow once again.
06:50Yeah. And generally speaking, when you first run this and kick this off, right, it's gonna ask you, do you wanna allow this? And especially if it's something that you're comfortable with, you do wanna read the what are you giving permissions to.
06:59But for the most part, if I'm using in it in and I wanted to be able to call it or execute it, just click always allow. Correct. Yeah.
07:05I know it goes back into that customized section where I showed, you know, you had the hand or the green check mark or block. So you can just change those settings once over there, or you can just change it directly here in this chat. But, obviously, you guys are doing this for the first time, so I didn't want to block that or always allow it right away.
07:22Anyways, it says the workflow ran successfully. Here's a quick summary as fetched.
07:26So this is a, uh, again, a collectibles video. We're talking about a 1977 Luke Skywalker card that sold for $687,000, which is crazy.
07:35The transcript over here is clean and ready. What else would you like to do with it? Blog posts, school post, YouTube titles, social content, or something else.
07:41Let's say, can you build out time stamps?
07:49Give me five for this video. And you can kinda stack this up, and you can have, you know, specific skills to build you transcripts.
07:57You could have a skill that essentially says, run this any end workflow. And then once you have the transcript back over here, turn that into the specific time stamps. And I have that on my other computer.
08:06But, obviously, for this demo, we'll just do it directly in here. Let's have this conversation. Great.
08:11And as you're doing that, I think one thing is important too is sometimes it hallucinates
08:15with time stamps.
08:17Do you do any, like, cross verifications or anything with it to, like, know that it's not just making things up? I sometimes will check the YouTube video on that. Um, to be honest, I use time stamps a lot more for kind of the keywords on YouTube and kinda ranking, um, but I'll just quickly double check.
08:31Okay. Here's the time stamp, what's on the YouTube video. But you can see over here, it talks about the sale at Heritage Auctions, comparing the two PSA tens that sold, other 1977 Luke Skywalker cards, Scanlan Australian cards, and then premium pricing and value comparisons.
08:45So I think this is good out of the gate. And this is, again, something that's kinda blocked in Cowork. You can't work with YouTube.
08:51And now we used n a n as that bridge where we can call an external API and then send the data back. Got it. And this one, currently, you're using Rapid API, which is more or less a market for APIs.
09:02Correct. And you can see also, just to show you, the execution log. So if we go over here to executions, a total of two seconds, it succeeded, and you can see all the green check marks.
09:11So if it fails, you can always obviously go directly in AAN and change things up. It shouldn't fail being super basic like this, but there's a lot more complicated workflows out there. Of course.
09:21Of course. So that makes a lot of sense. And then also in terms of that too, I imagine if you're trying to connect to Rapid API, you you might say, hey.
09:29I wanna connect to Rapid API. This is maybe some documentation around it or something like that to help you build out that HTTP
09:37node that's making that request. It can maybe give you some insights on actually how to connect that up. Yeah.
09:41You can actually build any end workflow from scratch. So you can write over here and say, hey. I want to connect this API.
09:47This is what I want in this specific workflow, and it will build this out for you. So I don't know if anyone remembers the community challenge about two months ago. Uh, they had it where you had to classify different emails and then use evaluations.
09:59I think it'd be kind of a pretty cool use case if we go over here and just start a very basic build. Obviously, we're not gonna go in full depth and detail of building out this workflow, uh, but maybe we can start with some sort of email classification.
10:11Yeah. That makes sense. Yeah.
10:12We we do monthly challenges for anybody that doesn't know around, like, different use cases with NAN, and that was the one that he was talking about from before. So, uh, if you wanna if you wanna try out some, uh, community challenges, that might be a good place to get started. Yeah.
10:25So let's try it. So I'm gonna say, can you create a brand new NAN workflow
10:29and call it I don't know. We'll say NAN community challenge.
10:34Since you have NAN already connected up to it, now because it has been connected, we can actually have Claude go ahead and create that workflow for us so you don't actually have to go into the actual workflow space in order to get this built out.
10:49You can simply ask it to do it, and it's gonna go ahead and create this for us. And what you're tying it paying right now, I'll just repeat, can you create a new in it in workflow? Call it in it in community challenge demo prod for podcast.
11:00Essentially, I want you to have a classifier that takes a look at the incoming emails, which we should stim simulate with a webhook and classify it into five categories, IT support, billing, feature requests,
11:15and I'm sure there are two more that are coming up here. We'll put spam, and then what one we'll throw one more over here. Let's see.
11:23Let's do I'll just change it to four.
11:29We'll keep it basic over here. And then I'm gonna say after we classify these.
11:36And then what I'm also gonna say, and I think this is really powerful within CoWork, if you are unsure, ask me any questions. So that's just one way that you could stop hallucinations.
11:48And I'm also gonna say, rate the initial prompts. I will tweak them later on.
11:54Alright. And let's give it a shot. So what it looks like you're using, uh, Sonnet in this as well.
11:58Do you when you build the workflow, do you typically use Sonnet, or you do Opus? Or I I use I use I use Opus on those workflows. I just didn't change it on here, which I apologize.
12:07Yeah. So it's all good. I just wanna make sure just for clarity wise.
12:10I yeah. I try to use I try to use Opus for creating the workflows as well. Um, but any of the low level, you know, write a prompt or something, I try to switch out the models to save on them the expensive token costs.
12:21Oh, a 100%. And especially, like, if you have a subscription to Claude, you know, whether it's a $20 plan or a $100 plan, uh, there's definitely ways to optimize it. And we're actually gonna talk about one of ways that any n can really help you out on there.
12:34But it'll probably take a few minutes for this to run. You can see right over here. Let me pull the s SDK reference and search for all the nodes I need at the same time.
12:41And then it says cloud wants to use getting SDK reference from any n. So once again, I'm gonna do always allow. We have another one.
12:48Cloud wants to use searching nodes from NAN. Always allow. Yeah.
12:51I do appreciate it, uh, describing what is the issues and what are you approving as you go through it. And I think one of the benefits of using In NN with this type of systems is that we generally wanna get a result. We wanna get some sort of output.
13:04Anytime we're making an automation workflow, we want a result. But there's kind of a trust but verify going on right now. So you wanna you want to trust the system, but the power of using N8N with these types of systems is it allows you to really verify.
13:18You can go into the back end of the system and know for certain by looking at the execution history and all of that data and say, oh, this is clearly doing what I wanna do as intended versus sometimes if you just do it straight with co work, you may not always know.
13:33And it might be doing things where you're trying to figure out why is this happening. So I I do like that use case where you you have this do the heavy lifting, but you can kind of you can be the ultimate verifier. Yeah.
13:42And, like, especially in other industries too.
13:44I am not sure about the medical industry, but I work in compliance. Like, you need to log every sort of interaction that you have with AI. So, um, just having a a specific skill in over here and asking it to do something might not be adequate enough for regulators.
13:59That's a bit yeah. Especially in the health space, removing, you know, PII or HIPAA compliance or any of those types of things, sir. It's it's very good to have a a a digital paper trail.
14:09We'll say that. Yes. And you guys have the guardrails node, which can help a lot on there too.
14:13For sure. For sure. Yeah.
14:16That's important with these things. So, yeah, I see it being built out right now.
14:21And so talk to me just a little bit about these patterns that you're seeing there. Can you talk to me as this thing's getting built out right now?
14:28Maybe a bit of the pattern. So let's just say you have a goal with building out some sort of automation, whether it's for yourself or for a client, right, and you wanna use co work and and it in together.
14:38What's your typical patterns for kinda getting started? How do you ideate? How do you get started?
14:43But tell me a little bit about your process. Yeah. So I think kinda my process on it, if I'm gonna build out a workflow, first, like, in the past before, I think Cowork and Claude got really good at building out workflows, I was very stubborn.
14:57I always built them out by hand. I was I'd figure out some sort of flow, and I'd go into any end. Okay.
15:02This node we should use here. This node we should use over here. But now I kinda look over here, and I say, what's the general idea of this workflow?
15:09Give some context over here. And, obviously, in a world scenario, I'd throw in a lot more context. Just being on the podcast, we're gonna keep it pretty short.
15:18But you throw in as much information as you want in the beginning and then just clean up the mess. I think one of the things that no matter what you're gonna have to do is rewrite prompts. You're gonna have to change some of the inputs and change some of the directions of the notes.
15:30That's given. But I think it could save you hours just starting out. And I think it's I think it's also really helpful for people that just haven't spent a ton of time in any end yet.
15:40If you wanna get your feet wet and you maybe you don't wanna look at a template and wanna start from something yourself, which I would still recommend starting with a template, you can go in co work, give it an idea, and then just see what it built, and then start taking a look at all the nodes. Um, again, very important that you have an understanding of how any end works.
15:57You just don't go over here and you start doing AI swap. Uh, it's very important to know. This is the node.
16:03This is what this node does. This is how you put this in order. This is how nodes specifically execute because I know for the first time when I was in NAN, I didn't understand that, and I built some, uh, janky workflows.
16:14But, yeah, pick that up. Janky is the word for it for sure. Yes.
16:18The the loop over items, I had no idea how it worked at first. I come from a co coding background, so I was like, oh, we have to always do it for everything. I didn't realize nodes went one at a time, so that was pretty fun.
16:30Mhmm. Yeah. And that's why you guys talk about it in the documentation.
16:33You might not need the loop over item node. Yeah. And as, uh, as you go through this, it's getting me thinking about, like, as these things are a multiplier.
16:43Right? So depending on your depth of knowledge, there's a lot of things that happens with these AIs that you just trust it. And the thing is with any of these LLMs, they say it with a lot of confidence.
16:53And unless you truly know how to cook in your kitchen with this type of stuff, you're just gonna agree with it. But sometimes it says things, and you're and if you actually know the back end of the system, like, it in, you can say, hey. Um, why are we doing it this way?
17:06Can't we do it this other way? You know, uh, instead of a code node, could we new use a dataset node or another node? And it goes, oh, you're right.
17:13Good catch. And then usually a good catch is like, oops. Sorry.
17:16I've messed up. Uh, let me do it another way. So it is it is it is good to do it not only as a a time saving tool, but if you do have depth of knowledge and this can teach you a bit as it builds, uh, I think, uh, it's very helpful to to to know the systems well.
17:32And I think it can also help you with debugging in some aspects. Like, imagine you built out a workflow two or three months ago. You don't remember all the nuances, and now you have some sort of bug.
17:40You can go back and forth and try to figure out, like, what is wrong. Maybe take a look at the execution history, grab some of the inputs over here, and use it for testing, especially if you have some sort of a webhook as the introduction to that workflow, or you could just add in a webhook and, uh, go back and forth. That's great.
17:56Yeah. And I know so you're connecting with the, uh, the the official n eight n MCP. Is that correct?
18:01Yes. What whatever you guys have on the customized tab. Cool.
18:05Yeah. And I just recently, uh, met with the, uh, one of the team members who actually created the n eight n MCP. And I actually found this out just earlier today is that you can actually use it to create data tables and and actually go through the process.
18:18Yeah. I just found that out too. And I was like, oh, wow.
18:20So I can actually make internal data tables and populate it with data. I was like, that's incredibly useful for me because there's a a lot of people use Google Spreadsheets as a way to as a kind of a I would call it a a ghetto, uh, data table, uh, database.
18:37You know? Uh, but Yep. Some limitations with credentials or if you make it as a template or if you're if you're trying to, like, share that template with someone else, now they need to hook up Google credentials.
18:47But if you're using data tables, you don't need to. Yeah. And that especially that happens in the enterprise side of things.
18:52Like, you're not gonna be able to get to log in to Google quite easily. One of the workarounds I do is I just have the form uploader, and then I'll just upload a spreadsheet and go through. Oh, that works too.
19:02That works too. It's a good move. Yeah.
19:04And these are all these little little little tips and tricks. And that's the thing about these spaces is that I think all of us tend to build with what we know. And then there's so many things happening so quickly in this space, uh, across all of the the AI platforms and everything else that, like, even someone like me and you are both like, oh, you can make data tables with this?
19:21I'm like, I didn't know that either. Like and so this is how the point of this podcast is to share these insights and best practices so people can get up and and iterate with this.
19:29I know. And I I will say that's one of the things I like with my full time job. So I I work at a company called Bill, and every Friday, they have different people that are building out AI workflows in any end.
19:38They kinda share what they built for the week and kinda their key findings. So there's probably, like, I don't know, 50 people on these type of calls, and they get to learn every single week someone else's viewpoint of building out automation. Amazing.
19:50And that's, like, that's the power of the community, man. Everyone's just learning and growing together and sharing insights and knowledge. Try not to get stuck in those, uh, private dev holes, you know, where you just Yes.
19:59You don't realize that someone else has got a solution. Okay. So I I think this is done.
20:03So you can see over here the workflow created despite 500 errors. The API timed out for some reason. Probably not a big deal.
20:09Uh, here's the structure that was built. So we have a email webhook, an email classifier. It says basic LM chain plus structured output parser.
20:16So interesting enough, it didn't use the text classifier node, um, which is interesting. Um, maybe we'll ask it to change that because, again, we talked about how you can go back and forth. We have a route by category, which uses a switch, which sounds correct.
20:29We have the the four different ones over here. Policy Google Docs, which obviously we have no Google Docs, but you could have a hypothetical Google Doc and probably just send it over here and say attach it. Uh, draft email responses with an AI agent.
20:41What I like also, it says before you can run it, you need to fill in three things. OpenAI credentials, which is interesting.
20:47It uses OpenAI in comparison to the cloud models. Weird.
20:50Right? Yeah. Uh, Google Docs, OAuth credentials, and then Google Doc URLs.
20:55So paste each policy URL into the respective node. To simulate the email testing, post it over here with a body like this, the classifier prompt. So now I'm gonna load up this in any end and actually see what happened.
21:06Alright. Let's load this up. Yeah.
21:08I saw someone too as you're getting that ready. Someone else was building with our own internal, um, in it and prompt system, uh, that we have inside of the the
21:16AI workflow builder.
21:17Oh, let's see what you got here. Look at that. Out of the gate.
21:20I mean, sure, I could build this out relatively fast. But, mean, I think about the time that it's saved. You could be doing some other task, and you already know you're starting with a webhook.
21:28I would change this over here, um, right now, and I know you guys just did new, uh, UI design, but I'd change this to the is this actually text class square? No. It's not.
21:37That's a basic LM chain. Apologies. I'll say it one more time.
21:40Three two one. I'd change this out to a text classifier, which maybe we can ask it to do so. We route by our category.
21:45We have the docs, and then we have our draft response. So I'm actually kinda curious what the prompts that it gave us. So didn't really give us a lot of context.
21:55I think this is not the best prompt, which is expected. And then draft email response.
22:02You can see it's passing in all the JSON, which, you know, for a lot of people, that takes quite a lot of time at first until you realize how it all works. I complete over here. Yeah.
22:13So let's actually go back. We let's imagine, hey. We don't want this over here, basic LM chain.
22:19This is bad. We should have used a text classifier.
22:21So I'm just gonna ask that. And as you type that in, that's the point of, like, if you know in it in well and you're like, okay. This is not the best use case or not the best node for this use case.
22:30Let's switch this out to something else that, you know, is a is a bit more, I'd say, focuses on proper design parameters.
22:39Right? How how do we do that? And I'm gonna say also there is no error handling,
22:45which, obviously, real world scenario, you'd want a error handling. Most tutorials aren't gonna issue error handling because it takes extra time, but we're gonna say we should throw in some sort of error handling. That's an important use case.
22:54Yeah. One of the good features with n eight n is that you can have these universal error handling. So if something happens, notify me either Slack or email or text message or whatnot so that you know.
23:04Because sometimes with automations, when it errors out and you don't know, when errors happen in automations, they they usually happen at scale. You know?
23:14So it's not just one error. You you might send an I I built a note or a workflow, uh, for somebody a little while ago, and I sent it through, and it aired out and ended up sending, like, 36 emails.
23:30And I was like I was like, oh, that's not it just just filled up the whole system. I was like, oh, that's that's automation
23:37messing up at scale. That sounds terrible. I added in actually one other thing as well, which we'll see if it picks it up because it's a little bit more advanced use case.
23:45I'm saying, we also add in a notice, send in an email, and wait for a response twenty four hours with the human in the loop within Gmail? Like, this is a specialized node within Gmail, uh, and we'll see if it picks it up or not.
23:56And, again, I'm using Sonnet. I should have used Opus. So it's not the smartest, but that's okay.
24:01Oh, we're already saving tokens. Yes. We'll talk about actually, do you wanna talk about saving tokens now?
24:07Yeah. Let's talk about it. Yeah.
24:09So as this is getting fixed, you know, one of the benefits I also see of using any end workflows is token savings. People, when we build out skills, they have a lot of different connectors, four or five. And, obviously, in cohort, you can chain together skills.
24:21So, like, when one skill finishes, you can start the second skill or the third skill, but this is gonna eat up your tokens a lot. And if you already have an any end subscription, whether at work or a personal $20 plan or $50 plan, you can just build out a any end workflow and kinda do what I had over here. Hey.
24:38Start this any end workflow. So instead of having all these skills run, all your tokens, and hitting those cloud limits, why don't you use the executions that NAN gives you, whether it's, like, 2,500 or 10,000, and, uh, run it on that platform?
24:51Yeah. Yeah. Yeah.
24:52Yeah. Leverage them. If you're smart with how to use your your tokens and leveraging systems that already exist, especially, like, if there's templates that are already out that exist, and then we I think we have, I don't 7,000, 9,000 templates in the template.
25:03Yes. Yeah. You can say, hey.
25:05Let's start with a with a with a good template as a base that someone's already solved this problem, and then it merely can iterate on that versus trying to create from the ground up. And you can just iterate in here. So you can see, like, I'm asking it to fix this workflow.
25:18So maybe you go on to any end's website, you find a template that you want to solve a situation. So instead of building out a skill, you have a skill that just says, call this any end workflow, and then you can process those results back over here. And I find that to be pretty good because you have one skill that calls any end workflow.
25:34The data comes back over here with respond to webhook, and then maybe you wanna take a look at something else. You have another skill that just analyzes whatever you have in here. For sure.
25:42For sure.
25:44Nice. And have you seen any good skills that that work well with NADN and and COWERTogether?
25:51Yeah. I on my other computer, unfortunately, have built some pretty cool skills, at least for my YouTube channel, because I always get super lazy with writing descriptions, time stamps, the keywords, and everything like that. So whenever I upload a new YouTube video, I just say I uploaded a new YouTube video, paste that in, and then it'll run it in a in workflow, builds out everything that needs, sends it back over here, and has it all formatted ready to go.
26:12And I could take it to another level and have it just log in to my account and, uh, populate everything if I wanted to. I'm kinda just lazy over there. Sounds bad to say on a podcast, but I should take that extra step.
26:22No. Automators are inherently aggressive aggressively lazy people. And what I mean by that is that you're like you're
26:28like, look. I understand that I could do this, and it would take me about thirty seconds each time, but I would rather spend thirty minutes so that I never have this problem ever again. And so there's, that aggressive to happen.
26:40Yeah. Yeah. Like, how do I solve this problem so I never pro have this problem again?
26:44For sure. Now now what I haven't looked at too much, and I think it would be actually quite interesting, is Cloud can control your computer. So different use cases of, you know, controls your computer first to grabs whatever you need and then starting any end workflows.
26:57I think that would be a a pretty killer response. So if anyone has any pretty cool automations, like, leave a comment on there. But if you didn't know, Cowork can control your computer.
27:05It can also control your browser. But as I mentioned with any end workflows, it is significantly better to just use some sort of scraper. Now if you're using, like, a popular social media site, whether it's, a LinkedIn or Instagram, thousands of scrapers, Appify,
27:19or other websites, as we mentioned a little bit earlier. If it's something custom, maybe you build out a Python scraper. Yeah.
27:25And that's things. Like, it can take over your, uh, system. It can use the browser, but it's really slow.
27:30Like, Correct. It is terribly slow. It eats through tokens.
27:34And, yeah, better off to build out some sort of any end workflow that can hook up to Appify for it. Yeah. Exactly.
27:41And and you might wanna do that as, like, a
27:43proof of concept or something that, oh, can you go in and do something on the the the Chrome browser or something? But but in in yeah. Because we always go with, okay.
27:53It might be out of the way. It might be, like, a big bloated system that takes a long time, but it's too long, especially, like, let's just say speed the lead.
28:01Right? If you're gonna have Claude all of a sudden, uh, go to a website, look up the information, grab the information, extract the information from a website, it's gonna be a lot slower than if you were to, uh, use some sort of scraping services that you could plug into that's, like, a dedicated code focused out the gate, solves a very niche problem.
28:20It's gonna be a lot faster. So there's kind of, like, trade off between kinda, you know, uh, can do anything but slow versus I've got a specific task. It can go fast.
28:29Or imagine you have a thousand pages that you wanna scrape all, like, different directories. How long would that take with, uh, Cloud right now versus the you either build out something custom in Python or you find something that's already online and you kinda hook everything together? Yeah.
28:43Yeah. Exactly. Yeah.
28:44So so leverage your power of in it in and scrapers if you really wanna build something that's kinda that's speed and reliable. I was talking about testing.
28:52Yeah. Trust, verify. I feel like that's the the the name of the game when it comes to this these AI systems is, like, you wanna trust it, but you also wanna verify it.
29:00And I think that's what the case is, that you feel really powerful when these AIs can do so much for you, uh, until you get burnt by it. And then you kind of learn to be a bit more wise with the verifying things work repeatedly.
29:12Uh, otherwise, uh, they will pop off and you least expect it. I was I was building something that would, like it would look at what my things were to do for the day, and then it started to break my thing into, like, chunks of time, five minute chunks, and put in my calendar.
29:24And I wasn't paying attention. And then also my entire calendar got filled up with, like, five minute slots of thing to do. And I was just like, I had to go through and just delete them all.
29:32And I was like, oh, man. So you you,
29:36you know, at the right place at the right time. Yep. And, obviously, like, with any end being structured, you know, you have the same exact workflow that should happen time after time.
29:44Obviously, you can split your workflow and go to multiple paths, but you have some sort of input. You have some sort of output. And if there's anything that you want to change, it's pretty visual.
29:52You know exactly where you wanna change something out. You can obviously go in over here and prompt it as I'm showing you guys right now. Um, but skills, not always the case.
30:00Obviously, you can change up skills. You can add the skills. You can remove skills.
30:04There's a lot of things that are really cool within CoWork. But if you want something that is structured, guaranteed to work every time, throw that any end workflow and just call it from CoWork. And then kind of on the testing side too, I know, obviously, we're building this out right now, uh, with this Gmail classifier gonna risk have a response back.
30:21You know, we can build out test data in CoWork as well. We could say, hey. Can we do five test runs in our any end workflow?
30:29And it'll essentially create a Gmail. It'll send it over as the webhook, and it'll go through that full workflow. Now I think this workflow will fail because we don't have any Google Docs attached to it.
30:39So I would hope that there's errors, and it would tell us this directly in here. But I would love to test it once this is ready.
30:45Sure. And while this thing is being loaded up here, I got a question. In terms of, uh, multitasking inside of here, do you ever spin up, uh, multiple agents, uh, different threads inside of coworker while you're working inside of n a n, dude?
30:58And, like, if you do, like, how do you break up those tasks so they don't step on each other's toes? I have. Yeah.
31:03So
31:04for example, I on my other YouTube channel, I do a monthly sales series of, like, top non sports cards, and probably most people watching have no idea or really care. But imagine I scrape eBay every single month, and I talk about the top sales for a specific category. Well, I used to make these PowerPoints by scratch.
31:20I would have to grab the images, download the images, write some descriptions, and everything like that. So that process now, I just fully automated with Cowork in any end. I just say, here, I wanna grab last month's data.
31:32I already have a scraper already built out. And then it'll go out there. It'll scrape everything for specific keywords.
31:37It'll go through, filter out the top 25. I have a human in loop aspect where I just do a quick review of the spreadsheet that's generated, give it a check mark, good to go, then it'll actually build out a full slide deck for me. I mean, this process used to take me three to five hours, and now it's done within probably twenty or thirty minutes of my time, which is really cool.
31:54But at the same time I'm doing that, I can build out another any end workflow, or I could have a conversation in here because Cowork kinda acts like a chat GPT. You can just ask random questions to an LM. It gives you the result responses for it.
32:06So I think of it kind of like as you guys had for ChatHub for a while back where you could essentially write to an LM and work with any end workflows. I just think this is the next level of ChatHub. Yeah.
32:17I can see that. Because the ChatHub makes it a simple interface that everybody understands, chat. And then it goes into the back end of automations, which not everybody understands.
32:24So being able to have a simple front end that makes a complex back end is is awesome. And and to add on to that, right, let's imagine you're in a corporate environment. You don't want to have everyone have access to your NAN workflows.
32:38So you could just essentially have your NA and workflows and have multiple people call into it based off of your settings that you set up and, uh, specifically customize. The issue with skills is skills live individually, I believe, on every single person's computer. Obviously, it's not audible and everything like that.
32:53But I think from, like, more of a security side of things and just getting accurate results every time, better just to call it any end workflow and then go back into Cloud Core. In addition, one thing I I think is worth noting, I think there's a lot of people that believe that anyone can use Cloud Code or can load up a terminal and understand things.
33:11They have not worked in a corporate environment. There's people at every single company that are brilliant people but struggle with technology, whether it's, like, on a Zoom call or using spreadsheets. But they know so much about their domain, and they expect people just to jump into Cloud Code and get started right away.
33:25I I don't believe that. But what I think using NaN and Cowork, like, Cowork is a chat interface. Someone can understand that.
33:32Like, okay. I'm just talking in kinda like a chat GPT. NaN is the same thing for people that wanna know how to code.
33:37Oh, okay. I have to use this node, it performs this action with an AI agent, and you structure it. So working in parallel, I think these are both ways for really nontechnical
33:46people to get their feet wet. And maybe one day, um, go into Cloud Code, but they don't necessarily have to. Mhmm.
33:53Yeah. It limits the barrier for getting up and getting started, and that's a hard thing. When a lot of people wanna start to automate things, there's this, like, frustration in the beginning that, like, you wanna get started, but then the the challenge is you hit these roadblocks and you're like, oh god.
34:07I don't understand it. But if you have a front end interface like this and you can you can access, well, why did this happen? What was the cause?
34:13What's this you know, you can you can learn about these things. They can help solve those problems, but you can learn from the systems
34:19without getting that typical thing where you just you just wanna build you just want you want a result and you want a quick win. And then you have these front end interfaces. You can you can get those results so much faster.
34:28Yeah. You can build your first AI agent in any end within twenty minutes. Now are you gonna understand all the nuances of any end workflows?
34:34Are you gonna understand proper prompting techniques? You know? No.
34:38That's not the case. But, you know, you can get someone super excited that they learned about AI automation, whether it's, like, someone pretty old that doesn't understand everything or someone that's a little bit younger. And I I really like it for that aspect.
34:49I just wanted to call out if you take a look at my screen right now, it says, Cloud wants to use updating workflow from any end. This is something that you might wanna choose only allow for this task. Because if you always allow this, it might make changes to your any end workflows that are already working perfectly.
35:02So a good call out of something that you just want for this task and just to manually review it. Obviously, I don't really care because this is a demo, so I'm just gonna click allow for this task. But really see what they mentioned because you can see 14 nodes over here.
35:15They're talking about some specific changes. We're gonna fly through that and say, hopefully, it's correct.
35:21Hopefully, that's correct. Yeah. That's the that's the trusting part that and I had it before because I had to build out a bunch of workflows, and I was telling it to fix some things and change some things, I had loud always loud.
35:31But then it would it would change my working workflows into broken workflows
35:36just because it was trying to fix another problem. What's also interesting, I think it's actually giving us a description for this workflow also, which is obviously best use case if you're setting up some sort of MCP connection. You should have a clear label for that workflow and a clear I think it's actually building out a description there too, uh, based off of these notes, but I'm not a 100% sure.
35:54Uh, regardless, what I do really like, take a look at this. What's changed? Remove the basic LM change, the the structure output parser and switch node, added a text classifier node, added error handling on five nodes over here, each error wired to a shared send error alert Gmail node, added Gmail send and await with a twenty four hour time out approve and decline buttons, added was approved branch, send a customer true, notify draft decline, added spam, notify spam received, Gmail so the team knows when spam is filtered.
36:23And then take a look at this because we we talked about, hey. Give us other suggestions. It set a confidence threshold, a one Google Doc credential.
36:30You have four separate credential placeholders for Google Docs. Yeah. We'll we'll ignore that for now.
36:34Uh, we'll actually talk about quick credentials in a second, which is gonna be really awesome. Uh, you can see a structured draft output and then clean up. You still have two duplicate workflows left from the creation retries.
36:44Worth deleting those from your any end canvas. So let's actually load up this workflow and see what happened. Great.
36:51Some good some good recommendations there. Yeah. We definitely wanna after this, we'll talk about the credentials connections because that's one thing that a lot of people get stuck on, especially if you're self hosting a lot of times because you have to go into, say, Google Cloud platform.
37:04You gotta make a lot of connections.
37:06Oh, so take a look at this. Now this is something that, you know, going back three to six months ago when people were saying, well, OMs can build out, uh, workflows in any end, I would get stuff like this. I'd be like, not quite.
37:17Um, you might get this the first time. Like, what the heck is happening? And, honestly, taking a look at this, I kinda get a little bit of that mess.
37:24Now Opus probably does a way better job. Again, I'm gonna throw that out over there. But this is an example.
37:28You have to know the basics of any n. So you're like, okay. This has our model, which I'm actually curious about model.
37:33It's using GPT four o mini. Yikes. Alright.
37:36No no worries. So we have this over here. Let's clean this up.
37:39And I know there's, uh, the button to clean things up also. Where do we have this? Right over here.
37:43Tidy up. If you guys don't know, click tidy up. Let's see what happens.
37:46Little bit cleaner. Alright. So I deleted that over here.
37:50Obviously, you can choose your own error handling as you specifically want. I know I was a little disappointed that there wasn't error handling after the webhook because after you have some sort of webhook, you wanna authenticate the webhook. Maybe if it's in the incorrect format, you'd wanna have some sort of error handling.
38:03So I was really hoping that that was the case. Obviously, we can go back to Cowork and prompt it to save you guys' times. I'm not gonna do that.
38:09Uh, we have our new text classifier over here, which I would assume the prompts aren't the best, but what's really nice is it grabs the from, the subject, the body as the category with the descriptions. Again, really like that. And if you're not sure, just click this over here, you can see a little bit more world scenario.
38:24You'd write way more thorough, but we're not worried about that today. We have our different documents over here. We have this draft email response.
38:32We have a review email response. So this should be the wait and respond back, and you can take a look, send and wait response. It says two placeholder value, which is funny.
38:41Obviously, we could just change this super fast. Says review draft. A customer support email is waiting for your review, and we have this over here.
38:48Now you can use this in way more services than just Gmail. I think probably the most popular one is Slack. It could be wrong.
38:54Or Telegram. And then you can see response type approval, approve or disapprove. So and we talked about twenty four hours.
39:00It built that out. So, again, think of the time saver. Like, oh, I have to build this out.
39:04I have to build this out. Like, we just had this running in the background as we were chatting. And then we have this was approved.
39:09Notify draft decline and then send to customer. Again, with all the JSON, I assume, already being in over here, which sometimes can be a pain for people, although it's super easy to drag and drop. Um, done.
39:19So, obviously, things I would tweak for a production build, but I think for a demo purpose, hopefully, you guys get the point of that. Yeah. Yeah.
39:26Yeah. It got most of the way there, which is pretty slick. And, again, we're not using the the most frontier model.
39:31It's okay. So we know things could get improved with that as well.
39:34Uh, this is this is great. So let's talk about these quick connections and how to connect credentials quickly. Absolutely.
39:41So in the past and we're gonna use Google as an example. You would have to go and go through a bunch of different documentation. You have to go Google Cloud.
39:48You'd have to figure out everything. You're like, man, this is such a pain. In fact, funny going back to my YouTube channel, some of the highest viewed any end videos are just credential setups, whether it's, like, Gmail or Google Sheets or Slack.
40:00But you guys just built out something that I think is really, really cool because it's simplified. If we go over here and choose credential, let's imagine we wanna create a new credential. You can see that you have a sign in with Google.
40:11So you can just go over here, sign in with Google, choose your account. I'll just choose this one. The others are blurred, and that's for my Ryan Nolan data account.
40:19I click continue. Well, I should say, I want to allow. Make sure you don't skip that step.
40:23Otherwise, you're gonna have a little bit of issues. Click continue, and take a look.
40:27Account connected. Yep. Five seconds.
40:30I remember my first time setting this up. I was like, man, there's so many different steps. This is super technical, and I'm a technical person.
40:37I was like, oh, this is a lot. And, honestly, anytime I'd have to set these up, I just have to remember the steps. Like, I I used to have a Google Doc.
40:43Okay. Step one. Step two.
40:45Step three. Or rewatch my old YouTube video. Now there's, like, no reason for anyone to watch that YouTube video how to set up Gmail credentials or any Google visit.
40:52Takes two seconds.
40:54Yeah. Yeah. It's it's amazing.
40:55Credentials are the biggest pain. So once you got that sorted, then you should be able to fly right through it, especially if you just go through and figure out any of the credentials you need to have set up, whether it's Google or name name the platform, and then you're you're good to go.
41:08Because then from then on, you just simply open up the nodes,
41:11select the credentials you want, carry on with your day. Yep. And you can see the credentials are already set up across the board.
41:16Like, I didn't even have to tell co work, hey. Set up these credentials.
41:20By default, any end, when you throw that node in, we'll set up that, uh, default credential, whatever you have in here over here. Uh, so, you know, we talked a little bit about this. Um, are there any other ways that people can save on tokens?
41:30Yeah. I mean, I think the the biggest one for most people that are in the any end ecosystem and then also use the cloud co work is to definitely have these any end workflows, especially as you know, if you're working during business hours, your tokens get burned through really fast.
41:45And I assume a lot of these automations are gonna be run during those business hours. So set up with some sort of skill, have that running any end workflow for whatever you want, and go back and forth.
41:54Obviously, there's a lot of different ways that you can save tokens within Cloud Cowork, but I think this is one that's really slept on in this industry for some reason. Got it.
42:03Alright. That's great. Um, and then we talked about I think we're gonna talk about a little bit about how execution histories are stored and logging.
42:10So, yeah, I think it's really important to take a look at these executions because maybe you have some sort of error and you need to, uh, deal with it, or you just need to see how often this is running. So you can just go over here to the executions. You can see this has run quite a bit.
42:23One thing I think is really helpful if you have an error, I always just click over here to copy to editor, and boom. We have those results.
42:30I think this helps with errors or if you're getting some sort of hallucination. You wanna see, hey. What data came in our webhook?
42:36Why do we have a hallucination a little bit later with a specific AI agent? Well, you can, uh, really see those specific details, and you can see our data is pinned over here in this webhook, and then you can see the green checkmark. Uh, we have the results over here, like, essentially pulled from that specific API.
42:52The other side of it, again, we talked a little bit about, you know, compliance or HIPAA. You need to log your data. Um, at least, like, if you're working with some sort of regulators, they need to see, like, what are you sending to an AI agent?
43:02What is the AI agent's response? Do are you guys utilizing any sort of human in the loop? If you're doing skills directly within Cloud CoWork, that's not always gonna happen.
43:10Whereas you build out a any end workflow, you can literally point it in the workflow. Like, yeah, this sends data into my database, and it is stored directly here. This is the input to that specific AI agent.
43:20This is the output, and it's all visual. You can explain every single node, every single action. Obviously, in Corework, you can explain, yeah, this is a skill.
43:27This is a prompt that is used every single time, but it's not structured. Whereas this over here and, again, we were just using three note example. The basic as basic as it gets, um, you see exactly what's happening.
43:39Yeah. That's a bit it's the visualness of the platform that you can see what's going on and understand why the errors are happening. It's amazing.
43:47So you had a hackathon at your office. And so what were the results of the hackathon?
43:52Yeah. So previously, before I worked at Bill, I worked at a company called Stacks, and, uh, there's a lot of AI frenzy last year.
43:58And one of the one of the things that was pretty cool is they actually hosted a hackathon with some of our customers that processed with us. There were different companies that came in, and the goal of the hackathon was to solve some sort of problem.
44:10One company had a problem with underwriting. Um, if you don't know, if you're doing any sort of payments, typically, a company will have underwriting to either approve or reject applications. So they had issues with their customers, um, not going through the full underwriting process.
44:24And one of the things they they figured out is, hey. Our IDs that we're getting from our customers, they just aren't in the correct format. And we would pen an application and send it back to them.
44:33So in this hackathon, this this group of people never used any n before. They saw my YouTube channel and kinda coached them through any n for two days.
44:42At the end of the second day, they were able to build out a pretty awesome workflow that they were able to apply to their business. So, essentially, what they did is they would take an input in for a webhook. They would have the input of a, uh, driver's license ID, and then it would go to a visual model.
44:57Okay. Is our driver's license not blurry? K.
45:00Good. Is the driver's license in color? Good.
45:02Because you get pended if it is in black and white. Um, and then you would use an OCR model to actually extract elements from over here. Okay.
45:10Does the pattern match to the specific state? Correct. Okay.
45:13Is the driver's license not expired? Correct. Okay.
45:16Great. Does the name match the application? Great.
45:19And there's a bunch of different data points that they would take a look at. I'm not gonna bore you guys with that. Um, but that was literally their first two days ever getting exposed to NAN, and they're able to build out something that they're gonna put back into their business.
45:30And I really think that shows the power of NAN.
45:33Yeah. It's and that thing is the beautiful thing about this is the ability to connect the people that maybe aren't as technical to the the problems that they're trying to solve. Because all ultimately, at the end of we're trying to solve problems.
45:45And so if you could say, hey. You don't need to be incredibly technical. If you're using co work plus n it in together, you can figure out and you you're bringing someone who their job is to solve a problem.
45:55Maybe they are verifying IDs. They can very quickly build a system, a solution, an automation that solves that problem because they're gonna know that problem intimately versus a hypertechnical person.
46:08They're gonna know the technical back end of systems, but they they're not gonna be as connected to the problem. So I like that closing the gap between the person who has a problem and the solution that's created. And whenever you can do that, it it there's so much more value because it's a it's a real world problem versus just making up solutions that don't actually have any problems.
46:26Yeah. And, you know, the sad thing is, like, people make fun of people if they can't get their Zoom camera correct or they have an issue with the spreadsheet. But sometimes they have so much domain knowledge, and if they could apply it to an automation, like, they would save a ton of hours.
46:38I think between that gap of co work and NAN, that can help solve that issue. So what you know, coming towards the end of the podcast here, what advice would you give to, uh, let's just say, a company that wants to use this type of technology? What would you tell them to get up and started and and and really get value from this?
46:55Yeah. So I guess my first question is, you know, how much experience do they have with automations? Do they already have any end workflows?
47:00Do they not have any end workflows? Are they just starting from scratch? Like, hey.
47:04We wanna build out our first AI agent today. Do they already have like, how far along is this company? So, yeah, so let's just say they have limited.
47:11They probably have, let's say, one or two people at part of the company that knows a lot about it. A lot of these other people are just people that work in different departments. Some work in accounting.
47:18Some work in sales. Some work in different places, but you want them to all have the power of AI and automation. They just they just they're eager, but they just don't have the technical chops.
47:29Sure. Absolutely. So what I would do with that two person team is I would start in one department.
47:34Let's just for technical stakes or for simplistic stakes, we're gonna say they're gonna go to the accounting department. First, I would kinda just kinda track what that accounting department is doing on a day to day basis. Because I think if you watch an employer, you know, a coworker, what they do, you might come up with ideas like, oh, we can automate this.
47:51We can save you, you know, five hours a week doing this. Um, so I'd first observe anything before taking any action. I know a lot of people are so gung ho.
47:58They're like, alright. We're gonna automate everything. But until you understand the context of why people do certain things, you're gonna mess up in the automation, and someone's gonna say, well, AI sucks.
48:06Well, it's really the person that implemented it. So I would first observe what they're doing. And then after maybe a day or two, come up to them with some specific ideas.
48:15Like, hey. I think I can automate this. Let's work on this together.
48:18And maybe that's where you kinda hold their hand and show them the process. Maybe you wanna start with this connection with NE and and Cowork and say, okay. Can you describe what you're doing?
48:27Write this out, and you start seeing the NE and workflow getting built on the side. Obviously, you have to figure out your credentials. You're gonna have to figure out some additional changes as you saw with the workflow that we built over here today.
48:38Um, but it kinda showed them in real time, like, oh, wow. This process took me five hours, and we just automated it really fast. Again, with world scenarios, you have to do a ton of testing.
48:47You're not gonna just build a workflow in an hour and wipe your hands and say, man, that's done. I just saved five hours a week. You have to throw an error handling.
48:53You should go through at least 50, a hundred, two hundred runs of a workflow, obviously, depending on what it's used for, to really confirm it. Um, but I would start with that and just kinda work department over department. Um, maybe your executive team can tell you or point you in the right direction what departments have a lot of manual
49:10processes, and you start over there. Makes sense. Yeah.
49:13So figure out what you're doing before you go anywhere. Then once you start once you start to identify those core problems, go, okay. What is the one, two, three top priority problems to try to solve that kind of has that it's kind of the the intersection between complex and value driven.
49:28Right? You don't want something super value driven that's super complex, but you also don't wanna automate something that's not gonna serve any purpose.
49:34So how do we find something that's fairly easy to automate that that also ideally solves a a core problem? And if you can find that and then work with some technical people on the other side so if you do get stuck, you can kinda hit the the panic button where somebody can come in and help you out. It makes sense that would get them up and running pretty easily without those technical barriers.
49:53Yeah. And that kinda goes back to the hackathon. Right?
49:55That one customer had issues with their applications getting pended. One of the biggest reasons was those specific IDs. Well, what was the reason?
50:02The IDs were in the incorrect format. So why don't we just build out a any end workflow that confirms the format of the ID? If there's any issues, it spits it back and says, we will not accept that specific ID.
50:12So now on their website, they just have a form. Someone submits the ID. It sorts out the NAN workflow, and then it can say your ID has been approved or your ID has been rejected.
50:21Think about that. In a day, you saved one of their biggest issues, which probably cost them tens of thousands of dollars. Absolutely.
50:26And that's thing. You just gotta find those use cases and integrate them into every piece of the business. Same thing with AI, I would say, is that, you know, everybody wants this all knowing AI that knows everything about you and your business, and we have automations that solve every single problem.
50:38But that's not how it works. You you get one small problem fixed at a time. You start to build out a bunch of different workflows, and then it stacks together.
50:44Then it feels like magic when you start having flows through the entire business. But that's that's how you need to do it, just kind of one piece at a time until it it builds into a a beautiful orchestration system. Yep.
50:54And I would also say don't always chase the newest technology. Like, new technologies can get adopted really fast and also dropped really fast. I think we saw it with, what, Cloudbot?
51:03Yeah. It got really hyped right away. And then, oh, man.
51:06We have a ton of security concerns. We're not gonna use this anymore, and no one's really talking about it anymore.
51:11I could be wrong. But regardless, if you can figure out an AI automation and it works correctly, you don't always have to change things.
51:17If something works, perfect.
51:19Exactly. Yeah. Yeah.
51:20Just want to solve the problem. Beautiful. Uh, Ryan, is there anything else you'd like to let us know before you tell people how to get ahold of you?
51:26No. Unless you wanna hear about that Goggins story. We kinda took a little bit of Just real quick, though.
51:30Yeah. Because, uh, you have a bit of a background with, uh, with, uh,
51:34running quite a bit. As as we've we've connected before, you've always been on a treadmill having conversations with me. Talk to me a little bit Goggins.
51:42Yeah. So funny enough, my first marathon, actually ran with David Goggins, which is so bizarre to say. I was running the Las Vegas marathon, and for some reason, I had the dumb idea of, hey.
51:52Maybe I can b q my first time, b q being a Boston qualifying time. At the time, it was a three hour marathon that you had to get. First half of the race, fantastic.
52:00I think I did, like, a one twenty eight. Second half, I started getting really bad cramps. I don't know if it was, like, being in the desert or I just didn't drink enough, like, liquid IV or whatever the case may be.
52:10I was cramping really, really bad. And kinda towards, like, the second half, towards the end, I think it was, like, mile 19 or 20, see this guy running past me. I'm like, that kinda looks like David Goggins.
52:20And someone points at him. He's gonna carry the boats, and he's, like, nodding. And yeah.
52:24And I'm like, oh my god. David Goggins just ran past me. I found out he was pacing his wife or girlfriend, and this was a week after he ran a race called Moab two forty, which don't know that race, 240 miles across Moab in Utah.
52:38Now a week after that, he's already running a marathon. Incredible. Anyways, I gotta run with him for a few miles, and he he said, this isn't my race.
52:45I can't finish it. I'm gone or whatever the case may be. And he left.
52:48And I'm like, no one is ever gonna believe my first marathon, I ran with David Goggins. I finished the race, and I'm walking, and I hear someone yelling at me saying, hey.
52:58I paced you those last few miles. Goggins literally waited at the end of the finish line. He's like, good work.
53:04You know, I'm trying to link up with everyone else that ran with us. Like, he could have easily said, man, I got spotted. I'm gonna dip.
53:11He literally waited and tried to talk to every single person that ran with him for a few miles. Ton of respect from him on that side of things, especially a week after running a 240 miler. When I ran my first 100 miler, I was, like, completely gone for a week and a half.
53:24I was oh, man. That was terrible. My feet were throbbing.
53:27I was so sore. And I even did a 75 miler two weeks ago, and I've already lost two toenails or not two weeks ago. Two months ago, I already lost two toenails and, uh, took a while to recover.
53:36I can't imagine 240 miles running a marathon next week. Then also just to tie it back, uh, is you had a someone part of your community actually made a a Goggins inspired workflow.
53:47Yes. Yeah. So I'm always on a treadmill on my calls.
53:50This is actually a rare occurrence. I'm not on a treadmill, funny enough. But, uh, because of that, he made a joke.
53:56Hey. You're like David Goggins, and I obviously told him that type of story. He's like, I have inspiration.
54:01This guy in our community, Ken Rogers, actually know him from the card space, and now he's in AI space also. Uh, he built out a full David Goggins workflow on there. I'm gonna ask him to actually upload it to the NAN website so then people can download it over there.
54:14I think that's a good call to action over there on that side of things. But he built it being inspired. It's like a telegram workflow.
54:19So every morning, it gives them inspiration like, hey. Make sure to get in your run or make sure to get in your workout. And it says it in David Goggins' tone, and that was one of his first workflows.
54:28Awesome. Yeah. You can use it for inspiration as well.
54:30It's incredible.
54:31Ryan, this has been awesome. If people wanna find your channel and find you, how do they do that? Absolutely.
54:36You can check out my channel on YouTube. It's Ryan and Matt data science. Obviously, as the name states, start as a data science channel, uh, kinda just as a way to build out a portfolio over there, and it's kinda drifted into more NAN and AI over the last few years.
54:49I really appreciate if you checked it out. We actually have a full seventeen hour NAN course, I think the longest one here on YouTube, and I think over a 100 NAN videos. So if you're not sure what a specific node does, obviously, you can ask inside Cowork, or you can watch probably a twenty or thirty minute video learning the nuances of that particular node and kind of the use cases for it.
55:08Um, you guys can also join our free school group. We have a call every single Wednesday. You can join in.
55:12You could walk on your treadmill, talk about NAN, AI, freelancing, does not matter. Uh, free group. A lot of people will charge you hundreds of dollars.
55:19It's free. Come join us. Incredible.
55:22Uh, Ryan, it's been an honor and pleasure, my friend. Much love, and I will see you on the other side. Really appreciate it.
55:27Thank you again for having me on. Take care. Bye now.
55:29Bye.
The Hook
The bait, then the rug-pull.
The cold open lands before the intro: sixty seconds, two clicks, and Claude is inside your n8n instance. The episode earns the title by showing exactly what the other 99% skip — the compliance edge, the token math, and the workflow corrections that separate a working build from a broken one.
How they asked for the click.
MENTIONED ON CAMERA
