15:57Mansel Scheffel · Tutorial
Claude Cowork Projects Just Changed Everything (AI OS)
A 16-minute screen-share tour of how to build a four-department AI operating system inside Claude Cowork Projects — no IDE required.
March 22ndThe argument in one line.
AI context goes stale the moment you stop feeding it, and a daily MCP pull from email, team chat, and call recordings is the only reliable way to keep an AI operating system's knowledge current.
Read if. Skip if.
READ IF YOU ARE…
- You use Claude Code or a Claude project for real work and have noticed it drifting — writing in a voice you stopped using or missing recent business decisions.
- You run a small business that already uses Gmail, Slack (or Teams), and an AI call-notetaker like Fathom or Fireflies.
- You are building a self-updating AI operating system and need the data-harvesting layer.
- You have done some AI setup but have not yet wired your existing tools into it.
SKIP IF…
- You have not set up a Claude project or AI operating system scaffold yet — this video assumes that foundation exists.
- You are looking for a no-code plug-and-play solution; the system requires MCP connector configuration and a scheduled task.
The full version, fast.
Most people configure AI context once and forget it — weeks later the AI writes in a stale voice, quotes old prices, and misses recent decisions. The fix is a three-source harvesting framework: Gmail extracts your writing voice in under three minutes, Slack delivers flagged knowledge updates, and Fathom synthesizes an objection library from your call recordings. All three routes land in a single local intake folder on a daily scheduled MCP pull. A refinement skill then reads the intake and proposes updates to voice profiles, SOPs, and skills — but a human must approve every proposal before anything changes, because auto-approving AI proposals can silently corrupt a sales playbook.
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 →Where the time goes.
00:00 – 02:06
01 · The two traps
Context set-once-and-forgotten vs. drowning in noise. The 'right slice' framing: too little starves the system, too much rots it.
02:06 – 02:56
02 · Pull, Push, or Export
Three mechanisms for getting data out of external systems. Pull via MCP is preferred; push via webhook fires in real time; export (cron/manual) is the last resort.
02:56 – 07:03
03 · Harvest 1: Voice from Gmail
Connect Gmail MCP read-only, scope to sent folder only, prompt Claude to distill a voice profile markdown file in under 3 minutes. Demonstrated live in Cowork.
07:03 – 10:40
04 · Harvest 2: Knowledge from Slack
Flag important Slack messages with emoji or 'save for later,' prompt Claude to pull flagged items via MCP, stash in markdown or Notion. Gate with sentiment analysis or human review.
10:40 – 13:05
05 · Harvest 3: Intelligence from calls (Fathom)
Pull Fathom/Fireflies call data via MCP. Synthesize an objection library, coaching takeaways (talk-time ratios, discovery vs. close), and product positioning intelligence.
13:05 – 16:06
06 · The daily sweep
One scheduled pull task (daily, end-of-day) pulling all three sources into a local intake folder. Demonstrated in both Cowork scheduled tasks and VS Code evidence/intake folder structure.
16:06 – 18:04
07 · Refinement chain and human sign-off
The intake folder feeds a skill that reads all flagged events and routes proposed changes to voice, SOPs, and skills into a proposals folder. Human reviews and approves each before anything is applied.
Lines worth screenshotting.
- AI context goes stale not because the system breaks, but because the world keeps changing and the context file does not.
- Too little context starves the system; too much rots it — the real job is finding the narrow slice that is both current and relevant.
- Gmail sent-folder analysis can reconstruct a complete writing voice profile in under three minutes via a single MCP prompt.
- Flagging Slack messages with an emoji or 'save for later' is enough signal for Claude to know which conversations deserve to be harvested.
- A pull reads a flag on schedule; a push fires instantly when the flag is set — choose based on how time-sensitive the data is.
- Call recordings contain three distinct intelligence layers: sales coaching data, objection maps, and product positioning feedback.
- Deals are won or lost at discovery, not on close — call analysis surfaces this pattern automatically once you have the pipeline.
- Auto-approving AI context proposals is dangerous: one misclassified Fathom note can silently corrupt a sales playbook and give an entire team wrong objection scripts.
- The intake folder is the universal landing zone — whether you use Cowork or VS Code, it is still just a folder on your computer.
- A scheduled task running a /data-intake skill at end-of-day is the minimum viable architecture for a self-updating AI operating system.
- Read-only MCP permission is all you need for harvesting — never grant write access to data sources at the intake stage.
- The human-in-the-loop gate is not optional polish; it is the mechanism that prevents compounding errors from propagating through the entire system.
Keep your AI current with three daily harvests.
WHAT TO LEARN
A one-time context setup has a half-life — the practical fix is a scheduled pull from email, team chat, and call recordings that keeps the AI's knowledge in sync with reality.
01The two traps
- Context quality is not about volume: too little leaves the AI uninformed, too much fills it with noise that degrades every response.
- The goal is the narrowest slice that keeps the AI current — not a comprehensive dump of everything you have ever done.
02Pull, Push, or Export
- Pull via MCP is the cleanest option because it lets you control exactly what gets harvested and when, without depending on a third-party trigger.
- Push (webhook) is worth setting up when data needs to land immediately — waiting until end-of-day for a time-sensitive signal defeats the purpose.
03Harvest 1: Voice from Gmail
- Gmail sent-folder analysis can extract a complete writing voice profile in under three minutes; the same prompt works for LinkedIn posts or YouTube transcripts if you want a channel-specific voice.
- Scope the MCP connector to read-only and limit it to specific labels or folders so sensitive emails never enter the pipeline.
04Harvest 2: Knowledge from Slack
- Flagging Slack messages with an emoji or saving them for later is enough signal for an AI to distinguish what matters from background noise in a busy channel.
- Stash harvested Slack content in markdown (not Notion) when it will feed a downstream skill refinement process, because the AI handles markdown more reliably than external database reads.
05Harvest 3: Intelligence from calls (Fathom)
- Call recordings contain three layers of intelligence: sales coaching patterns, objection maps, and product positioning feedback from real market conversations.
- Objection libraries built from call data are more accurate than manually curated ones because they reflect what prospects actually say, not what salespeople remember hearing.
06The daily sweep
- The intake folder is just a local directory — the insight is that everything, regardless of source, lands in the same place so a single downstream skill can process it consistently.
- A scheduled task running at end-of-day captures the full day's context before it gets buried; cloud-based routines are the fallback when the local machine is off.
07Refinement chain and human sign-off
- Never let AI proposals update a live system automatically — a misclassified call note can silently corrupt a sales playbook and give an entire team wrong objection scripts before anyone notices.
- The proposals folder pattern (AI writes suggestions, human approves before they are applied) is the minimum safe architecture for any system where bad data has downstream consequences.
Terms worth knowing.
- MCP (Model Context Protocol)
- A standardized connection layer that lets an AI assistant read from and act on external tools like Gmail, Slack, or Fathom without manual copy-paste. In this context it is used as a read-only data pipe.
- AI Operating System (AIOS)
- A personal or business configuration of an AI assistant that includes skills, voice profiles, SOPs, and context files — treated as a living system that must be maintained rather than a one-time setup.
- Voice card / Voice profile
- A markdown file summarizing how a specific person writes — greeting patterns, phrasing habits, tone, and dos/don'ts — used as persistent context so AI-generated drafts match their real style.
- Intake folder
- A local directory that serves as the landing zone for all harvested data (emails, Slack exports, call notes) before a refinement skill processes and routes the contents.
- Refinement skill
- An automated Claude skill that reads the intake folder, identifies what should change in existing context or SOPs, and writes proposals to a separate proposals folder for human review.
- Cowork
- The project-management UI within Claude that lets users organize tasks, context files, connectors, and scheduled tasks for a specific AI operating system project.
- Fathom / Fireflies
- AI-powered call-recording and note-taking tools that transcribe and summarize sales or business calls, producing structured data that can be harvested via MCP.
- Human gate / Human in the loop
- A mandatory review step where a human reads AI-generated proposals and explicitly approves or rejects each change before it is written into a live system.
Things they pointed at.
02:56toolCowork
03:20toolGmail MCP connector
07:03toolSlack MCP connector
10:40toolFathom
10:40toolFireflies
13:50toolVS Code
15:56toolClaude routines
09:20toolNotion
Lines you could clip.
01:00
“Too little starves the system, too much rots it — the job is the narrow middle.”
12:19
“Deals are won or lost at discovery, not on close.”
17:09
“You wouldn't want your systems getting updated automatically without you knowing what's in them.”
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.
metaphoranalogy
00:00So I'm sure you know by now that context is the most important thing you can feed Claude or any AI system out there. The bigger problem is is that most people set up their context once, and they just leave it thinking it's gonna work forever. But that's not the case.
00:12So what we need to do is figure out how we can keep our context constantly updating along with the things that we're doing every day. Now the great thing about context is that it already exists in most of the things that you've already been using even if you're not using some form of AI.
00:24Sure. You've sent an email in the last twenty five years, so that's one place we can harvest some information. Another one is Slack or Teams or whatever it is that your business is using.
00:33And then, of course, you are doing some form of call. Modern day calls, you're probably gonna have some kind of AI assistant. So for this video, I'm gonna be focusing on three ways that you can manipulate these in order to get certain parts of context and then later on update our systems.
00:45The important thing to take away here is that these things can easily be connected to just by MCP. We'll have a look at that in just a second. But before we get started, there are actually two parts to this problem.
00:53The first one being is that people aren't really doing this right now. They're not plugging everything in or thinking that they need to update their context. But then on the other side, you have people literally pulling in everything from every single system and just saying, yeah, man.
01:04This totally just refines my whole system for me in real time, and that's a lot of crap because all you're doing is filling up everything with bloat. So what we wanna do is try and find this little space in the middle here, this of just the right amount of information for the specific system that we're building.
01:18If we look at our little journey map over here, put out some previous videos, particularly on the discovery process of this. That was the data mapping and the pod mapping and the auditing. So if you haven't seen those videos, I'll link them down below.
01:29Particularly for this video, we're gonna be focusing on harvesting this information. So we've now connected all of our systems, and we're gonna work through these three examples that I spoke about taking out very specific information. This is very important for our refinement phase where we're either refining our skills in autopilot or refining our context inside our business operating procedures.
01:47Another really important thing here is that we don't want to pull in everything. Like I just said a few moments ago, we wanna have some form of quality threshold here, whether that is some form of grader that the AI is doing for us or whether it's you, the human gate. I definitely think certain systems absolutely need a human gate, and we'll get into that as a part of this prac.
02:04Now from my perspective, there are various ways you can get this information out of these systems that you're already using. One of them being a pull. So we can just use MCP, like I mentioned, to pull any information that we need out of these systems.
02:15Really straightforward to do, and Claude does most of the work for us. But that doesn't always work because sometimes Claude can't connect with an app or we need information sent at a very specific time, in which case we would push from that third party app to a central repo where Claude can access it or another form of process can access it to pull it into our system.
02:33You can think of this more as a reaction, and, generally, that's gonna be some kind of webhook that's doing the pushing for us. The third thing here, if these other two weren't available, would be some kind of export, whether that's you literally manually doing it or setting up some kind of cron job with programming that can export this thing for you out of the system that you're using.
02:49The point of what we're trying to do here is extract all of that information using one of these methods, preferably using this pull mechanism if it's available to us. So for our first example, what I want to do here is showcase how easily you can harvest your own voice just using your Gmails. For most of you, that's probably gonna be the best way to go.
03:06For those of you who've been writing with LinkedIn or maybe you've made some YouTube videos, you can obviously extract your posts or even your transcripts from your videos and get AI to analyze your voice that way. Realistically, using this process that we're about to use now, it's easier than it's ever been, especially if you take the Gmail route.
03:20And here we are in Cowork. You can see on the left hand side, I've got my project set up for my AI operating system. This is completely empty, you can pretend that we've just built this thing, and we're trying to get my voice for my sales outreach and things like that.
03:32So the first thing that I'm gonna wanna do is I want to head on over to customize, and then I want to go down to my connectors. Now for me, I've already got everything connected.
03:40And if you've watched my previous videos on data mapping, you will have done this too. But for those of you who haven't seen those yet, you could just come down here and you can connect all of the connectors that you would need where you're trying to pull out your voice from. For our examples here, we're obviously just going to use Gmail, and then later on, we're going to be using Fathom and Slack as well, which I've already got connected.
03:58The second thing that I want you to do once we are connected is make sure that you have not given this thing right permission for anything. All we're doing at this stage is reading. So as long as we've got read permission to go and extract what we need, that's a good start.
04:09Especially when it comes down to emails, you don't wanna be pulling in every single one of your emails because, obviously, there could be sensitive information in it. So something else I might do before this is I would create a specific folder where I want to extract my voice from. So that would be emails that I've sent to people so the AI can see how I respond to certain emails.
04:26I'd put anything that I don't mind sharing in those folders or maybe give it a specific label, and then I would just give this thing read access to read these specific user labels and go after those ones only. So that's step one of the process is to look at our sent only inbox. Then what we need to do is extract our style and dump it down into a voice card where it can distill all of the emails that it has and build a profile around our voice, specifically in a business context perhaps.
04:51If you were doing this from, say, a more personal profile and you were doing this for a life operating system, then you would obviously do it for personal emails. So then all we need to do is come back over to Claude, all we have to say to this thing is, can you connect to my Gmail m c p and map a profile of my voice by checking my last 20 sent emails?
05:09I want a complete profile written for this in a markdown file so that every time I want you to send me an email, you can reference that voice for business emails. And after about three minutes, this thing has found enough genuine samples that it can go and update its task and then start to write Mansell's voice profile. Now depending on where you are at the stage in your AI operating system journey, hopefully, you've been following along with my other videos where we build them from scratch inside Cowork.
05:32You can check that down below if you haven't seen them already. But depending on where you are in that journey, it's certainly gonna change how you use this. For this perspective, I'm doing it from the perspective of someone who hasn't built anything yet.
05:42We're trying to get those initial voice profiles as fast as possible, and you would do it as a part of your AIOS onboarding. And you can see how easy this was. Three minutes later, we have a one line summary of how I've been reaching out to people over the last 20 emails.
05:54The greeting is over here. Hi. First name.
05:56Always the name. Always a comma. Unknown recipient is always hey there.
06:00That's exactly right. The opening line, I always say, hope you're well. I hope all is well with you depending how well I know these people.
06:06If I'm pissed off with someone, it's usually short, and I will get immediately into the complaint without any form of greeting and things like that. So it has picked that out over here. It found some signature politeness and phrasing patterns.
06:16British, please can. Please can this be refunded? Please can someone from the team?
06:21Not can you please. That's an interesting distinction. Softened asked.
06:24I was wondering if I can. I thought it would be good too. I suspect there will be.
06:28I also use collaborative idioms, get the ball rolling intro with an enclosing offer before sign off. Please let me know if there's anything easier if you need from me. Point is here, in less than three minutes, everything has perfectly been mapped.
06:38Talks about my spelling, the dos and the don'ts of how I write these email, and it puts together some templates based on things that it's seen inside my inbox. So if you did this and you stashed it in this file, you would have a pretty good email writer moving forward. But like I said, that's just one of the pieces of mapping out your voice because there are totally different voices depending on the system that you're using.
06:56So you can just rinse and repeat for LinkedIn, YouTube, whatever it is that you do. Getting your voice has never been easier, especially if you've written something down before. But then there's something more valuable for this, and that's pulling out information that's updating old stale information.
07:09So let's say you built a system, and you had this perfect SOP which you turned into a skill. The problem is is that you guys have been chatting in Slack for ages about different problems that actually form a part of this SOP that you once wrote. So instead of having to go and manually update everything, wouldn't it be great if we could just get Claude to go and grab very specific context from the conversations that we've been having and pull it down into these SOPs and skills to then refine them later on.
07:34Well, spoiler alert, that's easier than it's ever been as well. So here we are in this demo Slack that I've got going. We've got a bunch of notes in here, and the way that I would think about this is, remember in the beginning of this video, I said to you that you don't just wanna pull everything into the system because then it would have to sift through all of this noise in here, and that's the last thing that we want.
07:51So let's imagine that I had more than just lonely old me inside this Slack channel, and a whole bunch of us have been talking inside a thread listing a whole bunch of bullet points. If we found out that this information was valuable, one of the things we could do to make sure that the AI grabbed this specific conversation is give it either some form of emoji or perhaps even save this for later.
08:12In doing so, we're giving Claude the ability to understand what's important to us. And then we can come down here, and we can say to this thing, can you please look at my Slack channel and tell me what I've saved for later so that we can pull it down into a system?
08:26What it's gonna do now is it's just gonna connect to Slack again. Like I said, we wanted to have this connected via MCP. It will then come into here and anything that we have saved for later.
08:34Let's let's just do another one quickly. Now we've got two saved. So it should pick these two things up.
08:39And so here, it found out two examples. It found our test, and then it found our AI intel. You can see that this is just some news that was sent to my Slack.
08:46The point is it's pulled out the things that we marked as important, and our next goal is to stash this somewhere. It's very important that we choose a place to stash it because we don't obviously wanna go through this thing manually every single time. We wanna build some kind of automated skill that runs on a schedule and pulls this out of our system at the end of every day once we've marked all the things that we deem important.
09:06For some of you, you might wanna stash it in markdown just like we did over here in our working folder. For others, you might wanna put it in Notion. It really doesn't matter as long as the AI can access it and it's in a system that's actually relevant to the work that is being done.
09:19For instance, if this was pulling down information related specifically to something you normally keep in Notion, just go and put it in Notion and then update the context accordingly. If this is gonna be something that is part of a later skill updating chain or something like that, you might wanna keep it in markdown because the AI will work better with that markdown as it goes through your skill refinement workflows.
09:38For me, I'm gonna select markdown, and we'll get into that towards the end of the video where I talk more about this whole chain and how it feeds all the way from intake through to refinement. A note on Slack, obviously, are different things you can use depending on the size of your business. If you have the Claw team plan, you can now use Claw directly inside here to literally go and do whatever you wanna do.
09:55So you never actually have to use co work. You can just use Slack. But for those of you who are using this to manipulate it as part of other systems, doing something via MCP is really a no brainer because it's the easiest way to pull out the context that you guys are working with every single day.
10:09And you can see here it's now stashed everything in markdown. Something else to note here, you might want to build some form of gating into this so that it has the sentiment analysis and the understanding of what's important to you.
10:20So even once you tag something in Slack, it can then intelligently understand what to do with that data. It's all part of building our system, and we'll get into that in just a second. Cool.
10:28So we've now figured out how to get our voice from Gmail. We've also figured out how to update our context from external systems. We use Slack as an example, but realistically, you can use Teams or anything where you're having conversations with people or systems are updating context outside of Claude itself.
10:43The third thing that we need to do now is look at another form of communication, and that is calls. So if you're using an AI notetaker, hopefully, you are because they're just life changing, you can use it in so many ways. It's not just for actionable summaries, but inside there is also training data for your employees.
10:58Say your sales team, for instance. If you wanna understand where their strengths are and where their weaknesses are, understanding what they talk about on this course is vital. More importantly, because you can help build further skills for them that help them write proposals later on, but also overcome any objections.
11:13If you analyzed 10 salesmen on your teams, if you looked at what people were saying on those calls and objecting to about your product, you could build an entire map of objections that people would understand how to get over before they've even jumped on the call. It can also feedback information about your product, the way that you're packaging it, things like that.
11:31We want all of that information pulled down for us. So, again, the easiest way to do that is to just use NCP. Okay.
11:37So I've come back over to Cowork now, and I've asked this thing just to create some dummy data for Fathom. I don't wanna actually go out there and pull out my calls on this thing because then I just have to blur the entire screen, which is no use to you. So all I've done is I've asked it to make some dummy data about the things we just spoke about, like the objection handling, the positioning of calls.
11:54But just know that what this thing would do is in the same sense of me asking for dummy data, I could have asked the exact same question asking it to pull that information out of Fathom because, again, we've just connected it via MCP. And here we go. This is the output that I got.
12:08Again, all of this is entirely realistic, you could do this right now with Fathom or Fireflies, whatever note taking app that you use. So we got a map on how our calls perform on average. It gives us all of the information and the read on what it gets from this data.
12:21It then tells us the headline coaching takeaways. So the deals are won or lost at discovery, not on closed. Bottom quartile reps pitch features before establishing the problem's cost.
12:31And that happens a lot with overly enthusiastic junior salespeople. It also analyzes the talk time and shows you where the correlation between who's talking and for how long influences your sales calls being closed. Second thing here, like I mentioned, is the objections.
12:44This is tons of data, not just for your salespeople to get smarter at selling, but also how you could potentially get feedback on your product before you've even sold it to someone to make it better and then make these objections completely mute. We then have our detailed objection cards that we could give to our salespeople ahead of a call after we've done client research and linked it to an entire workflow, really, and we get some product and positioning intelligence as well as a part of that.
13:07You get the point by now. We have this entire intelligence map. And what I want to do now that we have these three things in place is I want to show you what we can do with this data now that we actually have it.
13:18In its simplest form, the easiest thing to do is to have a schedule pool. We have a daily sweep over here, and what it's gonna do is we have our email, we have our Slack, and we have our calls, and we just want to pull in all that information as we need it. For the email, it might just be a one off thing, but particularly for the Slack and the calls, we would definitely want to be doing that thing every single day so that we're getting that fresh context and that our call context is updating to see if the improvements we made previously are working as our salespeople carry on down the line.
13:46And that landing zone can change depending on how you're working. For me, I work in Versus Code, so my landing zone is very different to those who are working in Cowork. While it looks different, underneath the hood, they're pretty much the same thing.
13:58It's literally just a folder on your computer. That's what we're trying to get away with here. We need to pull all the information down and store it locally somewhere where the AI can access it.
14:06So if you're working inside a coworker project, that would literally be in your context file over here, and you could just have any folder that you want. For me, I have a folder called context global, and that has various parts of the systems stored inside there that I use for this.
14:19To illustrate this more cleanly, we can look at what it looks like inside Versus Code. Again, this is just a folder structure on my computer, and you can see I've got a folder called evidence.
14:28Inside here, I have a folder called intake. Now the reason that I have this structure is because it ties in with my skill refinement process that I've covered in a separate video. But particularly here, the only thing we care about based on the things that we've been speaking about is this intake folder.
14:43So everything that I've been speaking about, you can see it ends up in here. We've got Slack. We've got CRM.
14:48We've got Fathom. It all gets brought into here. All of those raw events get pulled down and stashed in here.
14:54So, really, all that we need to do to make sure that this keeps happening is set up a scheduled task to do the exact same thing that we just did with coworker every single day at a specific time. And, again, we can do that quite easily just by clicking on scheduled over here inside our project. If you didn't wanna use a project, you could just click on that scheduled task tab there.
15:12But we can come over here, and then we can create a new task, and we'll call this context update. Give it a little description, updating context.
15:23Then you just need to give it some plain English commands about what you want. You can either do it this way, or you could build a skill, and then it's just a repeatable process anywhere that you run it. The point is as long as it's clear and you've got MCP connected, as you've seen, it can easily go into a system and get the thing that you want.
15:38For me, I prefer to use skills, so I would just do forward slash data intake, and it would run the skill at this exact time every single day. I can choose my frequency down here, and we could set this to daily. We could set it to a specific time probably towards the end of your working day after you've had a ton of conversations with people, perhaps 05:00.
15:56Something to note, your computer does need to be on if you're using a scheduled task. So if you did want to do this while your actual computer was off, if you were using cloud based solutions, you could just hook up Claude routines and have it run on a routine doing the exact same thing. So now you might be asking yourself, okay.
16:10Cool. So we've got everything in the intake folder. How the hell does the AI know to automatically go and update any of our context or update our skills and things like that?
16:18And that's why I made that separate video for that I will link down below. It goes into a very deep dive about how that system works from start to finish. So rather than reexplaining it, go and watch that one next.
16:28But at a very high level, it just uses a skill to go through this little evidence folder that we have over here. And once we have everything inside our intake, the skill runs through everything. It understands where context lives inside our system.
16:41So it knows where my voice is, where my business files live, and things like that. And then it cycles through all of those chats that we had in Slack and all of the calls that we had in Fathom, and it knows where to route any intelligent changes by making proposals that you can see in this little folder over here. I would then come in here as the human in the loop, and I would read those proposals and say, yes.
17:02Approve that change or no. Don't approve that change. It's very important to have that human in the loop.
17:06You wouldn't want your systems getting updated automatically without you knowing what's in them. For instance, this thing could think something is an objection from one of those fathom calls, and it might not be an objection at all, and then it's gone and updated your system completely incorrectly given your salespeople the wrong data, so on and so forth.
17:23You see the cascading problem here. So once we get to the proposal phase, I 100% recommend a human in the loop before you refine anything beyond that.
17:32But other than that, I hope I showed you how easily you can get this data from these external systems using MCP into your environment. However you wanna set that thing up, it really doesn't matter as long as the data ends up in a specific folder that you can tie in with other skills later on. Like I said, the entire data series will be linked down below so you can follow the journey for each video and get a very clear understanding of how to do this.
17:52If you have any questions, leave them down in the comments below, and I'll get back to you as soon as possible. Otherwise, check out the videos on the screen. They'll definitely help you in your journey, or you can check out my community where I'm helping AIS builders every single day.
18:02I'll see you guys later.
The Hook
The bait, then the rug-pull.
Every AI operating system has a half-life. You build it, load it with context, and six weeks later it's writing in a voice you stopped using, quoting prices that changed, and missing the deal you closed on Tuesday. This video is the fix.
Named ideas worth stealing.
02:06list
Pull / Push / Export
- Pull (MCP daily sweep)
- Push (webhook reaction)
- Export (cron or manual)
Three mechanisms for extracting data from external systems into a location Claude can access. Priority order: Pull first if available, Push if real-time is needed, Export as last resort.
Steal forAny system integration architecture where you need to decide how data flows from a source tool into an AI-accessible store.
01:24model
Harvest to Intake to Refine
The three-phase pipeline: harvest raw data from existing tools (Gmail, Slack, calls), land it in a neutral intake folder, then run a refinement skill that proposes context updates for human approval.
Steal forBuilding any self-improving AI operating system where inputs are noisy and updates must be curated before they affect downstream behavior.
01:00concept
The Right Slice
The principle that context quality is not about volume — too little leaves the AI uninformed, too much fills it with noise. The right slice is the minimum current-and-relevant data for the specific system being built.
Steal forDeciding what to include in any AI context file or knowledge base.
How they asked for the click.
VERBAL ASK
17:19next-video
“Check out the videos on the screen. They will definitely help you in your journey, or you can check out my community where I am helping AI builders every single day.”
Standard end-card with community plug (skool.com/ainative). No hard sell. References prior videos in the series for continuity.
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
Visual structure at a glance.
