I Have 3 Kids and am Doing Another Freaking Startup. So I Built an Anti-Forgetting Tool. (It Mainly Works, Too.)

I was tired of being the guy who builds systems for a living, but has so many plates spinning I couldn't remember to follow up on a damn email. Three kids (11, 8, and 5), a new startup, and the operational and life-dealing overhead are winning. So I built something to fight back. The whole thing is a single text file, about 510 lines of instructions that connect six business systems and fire two dozen automated triggers. It's not elegant. But it catches the things I drop before they hit the ground (most of the time).

Too many tools, not enough glue

I've spent the better part of two decades drowning in tools and processes. If you've been anywhere near the back end of scaling EdTech/(any tech) companies, you know the drill. Every tool helps, but every tool is also one more thing to manage and context-switch into. Another login. Another dashboard. Another place where critical information sits quietly until you forget it exists (or, worse, until a client reminds you it exists).

The dream is always the master control dashboard. If you're an IC it kind of works already. HubSpot, Salesforce, your PM tool, whatever. Everything flows into one place. I have been leading startups and scale-ups for years, so my context has always run across the business. It had the data, but there was too much of it. It could tell me what happened, but not what mattered right now, and it forgot (actually, that was my brain dealing with too many dashboards). Nothing telling me "here's what needs your attention this morning." (Technically not true, firefighting always needs your attention this morning)

I cofounded 8Levers with Lori Niles-Hofmann at the end of November. Bootstrap advisory firm built on Lori's Eight Levers of EdTech Transformation framework, helping learning organizations and EdTechs figure out how to scale their learning functions without burning everything down (or helping them burn it down to rebuild). Small team, running hard, and the plate-spinning was already becoming annoying in a very familiar pattern.

Around the end of January, I started reading genuinely interesting things deep in Reddit and Substacks about what people were building with AI assistants. Not the "10x your productivity" listicle crowd, or “OpenClaw changed my life, here's how it can change yours” Medium/LI posts. People actually building things, sharing what broke and what didn't. I had a real problem I wanted to solve, and the best way I've ever learned anything is by getting dirty and breaking things until they work (25 years of preaching experiential constructionist-based learning, and I'd be a hypocrite not to practice it).

My business problem. Can I use Claude CoWork to organize work-related tasks/processes/flows so I can free up brain space to focus on life missions outside of work without going full grey? I'm going to value that at 10x time-to-revenue returns (just in case any VCs want to put some money in). I gave myself until April 1st to run the experiment (as of writing, I still have some time).

What I actually built (with my buddy CoWork, also known as JoshBot9000)

The instruction file sits between the six systems I use to manage my day-to-day for work. Email, both calendars (work and personal, because doctor appointments don't care about your project pipeline), call transcription, CRM, shared drive, and CoWork. Instead of tabbing between them separately, JoshBot9000 reads from all of them at once and cross-references what it finds (CoWork being the one ring to rule them all).

Here's what the architecture looks like (Thanks, Claude, Graphic design skill I <3 You):

When I kick off a session by saying “good morning,” JoshBot9000 reads my inbox, checks both calendars for conflicts, scans for new call transcripts, pulls and processes them, checks the project pipeline for anything going stale, and hands me a prioritized view of the day. Not "here's everything that happened since yesterday." More like "here are the three things that will bite you if you ignore them, and here's what you need to talk about for your 10:30 call." It knows my 4:15 is non-negotiable because I'm picking up my kids. It knows which projects have been sitting too long without movement. It knows if I haven't published content for 2 weeks (or 4 weeks, in the case of this post).

It is becoming a very good personal AI assistant who has legit Bo-Staff skills, but sometimes lies to you (basically, my five-year-old with crazy tech chops who watches Napoleon Dynamite on repeat).

The core of the whole thing is a trigger table. About two dozen conditions (and growing) that fire automatically without me having to remember anything (other than making sure everything works at least once a week).

Auto pull a call transcript from Otter using an MCP connector? JoshBot9000 processes the whole thing. Structured record - commitments extracted - CRM updated - relevant intelligence flagged - and a heads-up if anything I just heard should change how I'm approaching that project. Got a meeting in two hours? It pulls the person's file, our last conversation, what I promised last time, where things stand, and whether anything relevant came through email since we last spoke (deepest apologies if I still haven't responded to your email. That is me being a shit, not my buddy j-bot9k).

It's Friday, and I haven't done a weekly review. It prompts a 10-point checklist covering task velocity, research cadence, project pipeline, and capacity planning. I can't dodge it (I've tried, and yes, if it were doing my quarterly review, I would need improvement: the jerk).

We also built a file-based memory system. Over 250 files (and growing) organized by person, company, call record, research, and projects. If I mention someone's name in conversation, JoshBot9000 checks their file and surfaces context on my next interaction. "You last spoke on March 3rd, they mentioned a restructuring of their L&D function, and you committed to sending the framework overview by the end of the week." The stuff I used to keep in my head and occasionally (to be honest, frequently) drop.

But the part I rely on most is the part that catches things before I forget them. Every business has things that fall through cracks because nobody remembered to check. Projects go stale because you are too busy herding cats. Invoices sit because the close-out process stopped at "deliverable sent." Content doesn't get published because there's always something more urgent (and there is always something more urgent). The trigger table catches those. It monitors thresholds and surfaces problems before they become fires. Internal project deliverable overdue? Flag (with some programmed cursing). Invoice due in 3 days? Flag. No content published in 14 days? Flag. Loudly (with some choice language).

What it showed me about myself

I have always been good at building systems for others. Building them for myself has historically been an organic hodgepodge of strategy with mission-critical triaging as priorities change. I know how to build on mission, but when you are usually less than 10 people (story of most of my career), you spin all the plates.

When you set up a trigger that flags internal strategy projects sitting untouched for 10 days, you discover you had things going three weeks cold and didn't notice (hello new website). When you build a content cadence tracker, you realize you hadn't published in a month (quit reminding me, I'm doing it now). When you wire up the email sync and cross-reference it against your CRM, you find that a prospect responded five days ago, but you never saw it because it got buried in a thread.

The JoshBot doesn't just catch the balls I've been dropping (or letting slide). It showed me how many I was already dropping without knowing it. That part stings a little, but it also allowed me to start to triage my system, and hard.

And it keeps stinging, because the system doesn't forget. The thing is, it doesn't have bad weeks (well, it is an AI, so sometimes I feel like it has been doing some LSD). It doesn't get pulled into a fire and forget about the other fire. It just sits there keeping score, and when you haven't done the thing, it tells you. Every single time. It's like having a business partner with a perfect memory (ish) and absolutely no interest in sparing your feelings (or in my case, two of them, JK Lori).

Is it perfect? Hell no.

When you work with AI every day, you always have it in the back of your mind that it could be lying to you. Or being overly and annoyingly polite. Or offering up meaningless flattery that, as a human, you appreciate in the moment but can sometimes forget isn't actually true. It's permanent gaslighting without the intentional meanness. It will confidently present fabricated information as fact. Just straight up invent things with a poker face. And it will tell you your mediocre idea is brilliant if you don't build something that stops it from doing that. I don't really like the idea of my AI taking my mom’s job with me (telling me I'm brilliant is a very important role).

The failures were frequent, especially early on. My email sync protocol missed a response for almost a week because the time window was hardcoded to 24 hours, and I'd skipped a couple of sessions over a weekend. I didn’t catch it until I manually checked my inbox and found an important conversation sitting there unprocessed. The trigger that was supposed to catch stale projects? First version fired on everything, including projects that were intentionally paused, so I ignored the alerts, and when a real one went stale, I missed it anyway. The call prep system once surfaced a forwarded newsletter as the "most recent interaction" with a prospect because it couldn't distinguish between a direct email and a mass forward. I walked into that call with the completely wrong context (lessons learned the hard way are the best lessons, according to the old GI Joe cartoons).

Every one of those broke something. And everyone taught me something about how annoying AI actually is and how to use it better in a work context (not getting a company salary keeps you really, really focused on what matters). Widen the email window to cover gaps between sessions. Add a "paused" state to project tracking so the alert fires only on items that should be moving (but also remind you to check back in and assess how things are going). Filter the email cross-reference to exclude mass sends (also to look through newsletters and see if there is anything relevant to what we are working on). The system gets better because when it breaks, I fix it. It does break. Regularly.

I built audit checks because I flat-out don't trust the system. Every session logs what happened, what got missed, what fired, and what didn't. A health scan runs at startup to ensure file counts match, indexes are up to date, and nothing has drifted while I wasn't looking. A consolidation protocol reviews the last five sessions for recurring patterns, because if the same thing keeps slipping, the system needs fixing, not just flagging. If the same protocol fails to fire after three sessions running, it surfaces the problem with a diagnosis: is the rule broken, stale, missing, or just contextual?

I recently discovered that security is an issue, and blindly allowing your JoshBot9000 to pull emails could expose you to all sorts of interesting prompt-ingestion attacks, so I did some research and put some firewalls around sensitive areas.

For anything that could end up in front of a client or in published content, there's a research verification protocol. Every citation gets traced back to its original source. Vendor-funded research gets flagged as commercially motivated. Round numbers without methodology get flagged. So do "future of work" predictions that aren't verifiable. If I'm putting my name on something, the data behind it better be real (and yes, I do read the source material, can't outsource that to JoshBot9000). Are these checks bulletproof? No. But they keep me and the system honest and auditable (and force me to follow up on a New Year's resolution to read more), and that's the minimum bar for trusting any of this. I use it for collection, cross-referencing, and surfacing. I never let it auto-generate anything a client sees without my hands on it first, though it has gotten fairly good at quality control and pushing back.

The time math

I want to be upfront about this because the AI productivity conversation is drowning in inflated claims, and here I am adding to that pile!

In the first two weeks, I lost more time than I saved. Debugging triggers that fired wrong, fixing rules that didn't account for edge cases, learning what Claude could and couldn't actually do (that last one was so damn annoying – what do you mean you can't magic AI it?). The rabbit hole I went down because learning about this was so interesting (it was a new toy and I am hooked) that I'm seriously happy I invested the time.

Before this system, operational overhead consumed somewhere around 15-20 hours a week. Email triage, CRM maintenance, call prep, call follow-up, pipeline tracking, company strategy research, research and reviewing the research. Not because I was bad at it, but because properly maintaining your systems as a small team just takes that long. Most people don't do it properly (honestly, most people can't afford to), which is exactly why things fall through cracks.

Now it's roughly 3-5 hours of review and decision-making (more if my spidey senses are tingling at some AI BS). I'm not burning 20-30 minutes before each call digging through four different systems for context, or another 20 after the call updating everything by hand. Conservative estimate: 10-12 hours per week back. Over a month, that's roughly a full work week I get to spend on actual consulting, building our business strategy, writing, and relationship work instead of feeding the admin machine (I usually spend that extra time with the family).

How I actually started (and how you can too, if you haven't already produced something much cooler)

It's not finished. I don't think it ever will be. Right now, it only works for me, on my desktop, in my sessions. Other people on the team can't touch it yet (I tried building an online version and correctly got pulled back by JoshBot9000 as a non-mission-critical time suck). Yes, I'm now learning about Neo4j graph databases, and how puppygraph can add whole new layers of coolness to the system (next vacation, I'll see what I can do there).

But if you're reading this and thinking, "I don't even know where to start on something like this," don't worry; it's easy. Start with something that really annoys you. The 510-line version (now with 37 subroutines) didn't start at 510 lines. It started at about 30, and a conversation with CoWork about how to automate things.

My first rule was a calendar sync. One instruction: read both my calendars at the start of a session and flag conflicts. Took about 30 seconds to write. Caught a double-booking the next morning that would have made for an awkward email to one of the parties or a very disappointed kiddo.

The second was a call prep trigger. Before any meeting, pull the contact from HubSpot, check for recent emails, and surface what I know about the person and where things stand. That replaced 20 minutes of manual digging across four tabs before every call.

The third was an Otter integration (or call recorder of choice). When a call transcript lands, process it automatically: extract commitments, update the CRM, create follow-up tasks, and schedule the next meeting or flag it as a to-do. That one alone closed the gap between "good conversation" and "nothing happened afterwards."

Three rules. Maybe an hour total to write and test. That was the entire system for the first week. Each one solved a specific problem that was costing me real time, dropping real balls, or was just another set of eyes on something. Then I added another rule. Then another. Then I noticed patterns in what I was forgetting and built triggers for those, too. Six weeks later, 510 lines, 37 subroutines, 29 automated triggers, 8 new skills.

The hardest part wasn't the technical build (CoWork is a great CoPilot). It was trusting the system enough to stop doing things the way I've always done them (and setting up my work folder on Dropbox in case I totally messed something up). I kept catching myself checking HubSpot directly instead of waiting for the morning briefing, or manually prepping for calls I knew JoshBot9000 would prep for me. Months of muscle memory screaming "if you don't do it yourself, it won't get done" (to be fair, that instinct kept me alive for 25 years before this). That's the real transition. Not learning the tools. Unlearning the habits.

And yes, I realize the irony. I help organizations build learning operating systems for a living. The exact problem I was solving for myself (nothing connected, no triage process, no system making the whole thing work together, and not aligning this to business problems) is the same problem I see in enterprise L&D functions every week. I knew I had the problem. I just kept putting it in the "do it tomorrow" pile until tomorrow was costing me real money and real relationships.

The system that fixed it started at 30 lines and cost $20 a month. Six weeks later, it helps me stay on top of my chaos. It's messy, it's opinionated, and it breaks regularly. It's also one of the most useful things I've built for myself in a good while. I love the learning curve. Now, please, someone share something I don't have to maintain 😛

So here's the whole thing: In the comments section, if you type "Cortex OS," Skynet will come online, and Sarah Connor will go back immediately. Just kidding dont do that (we don't need Skynet in our lives). I'll release it eventually, but I need another month or two of internal testing before it goes into the wild. Happy to talk about how to get started, though.

I apologize for such a long post, but I didn't have time to make it shorter 😛

~Blaise Pascal