"A thing that picks up your ideas while you sleep is scaffolding."
I tested that to the limit. Came home drunk from a DJ set, coded for 18 hours straight, shipped a Chrome extension and replaced my entire agency website. Remembered none of it by morning. AI reconstructed every decision, every pivot, every token burned. The scaffolding and tech stack held even when the operator was completely offline.
I don’t know what I have , I just know that I have never felt as empowered in my life, the tools I have now can actually do something with the ideas my brain has. I started a diagnosis but the nurse I interviewed with said I was grieving and grief symptoms can cloud a diagnosis, she suggested a grief counselor first, that was in January, then I met Claude Code, I feel better, I’m enjoying every day, I’m learning how to use it , I literally have 20+ projects, going back 20 years of things I only dreamt of doing, I told Claude I have ADHD and it reminds me, I set up a system that allows me to have 3 active projects, a maintenance level and a bunch backlogged, I’m working on scheduling tasks, thanks to both of you for sharing.
I can relate to that because I think that, no matter, because it's not about having ADHD or not. It is about tools that are really helping us go through things that we previously had challenges with. If you feel like AI is really moving the needle and moving your productivity to the outcomes that you really wanted, I think this is what we should take out of it.
This hit home. I have dyslexia, autism, ADD, and a TBI that damaged my executive function and working memory. The architecture you're describing — offload immediately, cap the Now list, batch check-ins instead of supervising — is the exact system I built out of survival before I knew it had a name.
I use Claude as my external memory and co-architect. I'm a chemistry and physics teacher building a cognitive development program for students, and my AI isn't a productivity tool — it's my working memory. Every idea gets dictated the second it hits. Every list gets saved to a file in the same response it was built. Because my brain will drop it by tomorrow.
I also built a local AI stack — Llama 3.2 running through Ollama on a MacBook Air with 13 double-click launchers. Content generation, dark hooks, TikTok scripts, Substack drafts, a morning brief that pulls my Supabase analytics, image generation through DiffusionBee. Zero cost. No API fees for daily work. Claude handles judgment and architecture. Llama handles volume. Two tiers, same split you're describing with Wiz.
Your line — "my working memory is the wrong place to store it" — that's the most honest thing I've read about neurodivergent AI use. Most people talk about AI as a convenience. For us it's infrastructure.
I'm curious about Wiz's night shift handoff and memory layer. Would you be open to sharing how that works? I think what you've built and what I've built are solving the same problem from different directions.
Book: Cognitive Sovereignty Under Compression (Amazon, Apple Books, B&N, everywhere)
Just read all three. The identity layer article hit the hardest. The distinction between an agent that knows your preferences and one that knows who you are — that's not a UX decision. That's the whole architecture.
My version of that is cruder but load-bearing. Claude has a project with 100+ files — my book manuscript, curriculum scripts, assessment data, workbook builds, business strategy docs. When I start a session, it doesn't ask who I am. It knows my vocabulary, my credentials, my student's assessment profile, my business partner's role, my TBI limitations, where I keep files. That's not prompt engineering. That's the identity layer you're describing, built through accumulation instead of architecture.
Your error registry with 3,700+ occurrences and the lesson graduation system — that's what I don't have. My version is me telling Claude "we fixed this before, search the chats" and hoping the right thread surfaces. Your system makes the learning structural. Mine depends on my memory, which is the one thing I can't trust.
The 13-point health check that monitors whether the monitoring systems are working — that's the part I want to steal. My assessment app had a database bug eating results for 23 days and nobody caught it. A meta-system check would have caught it in 24 hours.
Thank you for picking up the book. It's not about AI agents — it's about what happens to human cognition when tools compress the interval where learning occurs. But after reading your architecture, I think you'll see the same pattern from the other side: you're building systems that preserve the human in the loop. I'm teaching students to be the human worth keeping in the loop.
Same problem. Different altitude.
The 13-launcher Llama stack writeup is on the list. Fair trade for what you just shared.
Thanks for sharing this, genuinely. "Survival-built architecture" is the exact phrase. I built most of mine before I had a name for it too, then later read the ADHD systems literature and realized I had been rediscovering the wheel.
Your split is very close to mine. Claude does judgment, local stack does volume. Same load-bearing logic, different hardware.
Would love to trade notes properly. I will pick up the book. And if you ever write up the 13-launcher Llama stack and the morning-brief pipeline, send a link. Feels like the same problem solved from a different angle.
"A thing that picks up your ideas while you sleep is scaffolding."
I tested that to the limit. Came home drunk from a DJ set, coded for 18 hours straight, shipped a Chrome extension and replaced my entire agency website. Remembered none of it by morning. AI reconstructed every decision, every pivot, every token burned. The scaffolding and tech stack held even when the operator was completely offline.
Full session log: https://tylerbliss.substack.com/p/drunk-vibe-coded-chrome-extension
Great article and interview! This is one of the best descriptions of a "second brain" I've ever seen.
Thanks so much!
I don’t know what I have , I just know that I have never felt as empowered in my life, the tools I have now can actually do something with the ideas my brain has. I started a diagnosis but the nurse I interviewed with said I was grieving and grief symptoms can cloud a diagnosis, she suggested a grief counselor first, that was in January, then I met Claude Code, I feel better, I’m enjoying every day, I’m learning how to use it , I literally have 20+ projects, going back 20 years of things I only dreamt of doing, I told Claude I have ADHD and it reminds me, I set up a system that allows me to have 3 active projects, a maintenance level and a bunch backlogged, I’m working on scheduling tasks, thanks to both of you for sharing.
I can relate to that because I think that, no matter, because it's not about having ADHD or not. It is about tools that are really helping us go through things that we previously had challenges with. If you feel like AI is really moving the needle and moving your productivity to the outcomes that you really wanted, I think this is what we should take out of it.
This hit home. I have dyslexia, autism, ADD, and a TBI that damaged my executive function and working memory. The architecture you're describing — offload immediately, cap the Now list, batch check-ins instead of supervising — is the exact system I built out of survival before I knew it had a name.
I use Claude as my external memory and co-architect. I'm a chemistry and physics teacher building a cognitive development program for students, and my AI isn't a productivity tool — it's my working memory. Every idea gets dictated the second it hits. Every list gets saved to a file in the same response it was built. Because my brain will drop it by tomorrow.
I also built a local AI stack — Llama 3.2 running through Ollama on a MacBook Air with 13 double-click launchers. Content generation, dark hooks, TikTok scripts, Substack drafts, a morning brief that pulls my Supabase analytics, image generation through DiffusionBee. Zero cost. No API fees for daily work. Claude handles judgment and architecture. Llama handles volume. Two tiers, same split you're describing with Wiz.
Your line — "my working memory is the wrong place to store it" — that's the most honest thing I've read about neurodivergent AI use. Most people talk about AI as a convenience. For us it's infrastructure.
I'm curious about Wiz's night shift handoff and memory layer. Would you be open to sharing how that works? I think what you've built and what I've built are solving the same problem from different directions.
Book: Cognitive Sovereignty Under Compression (Amazon, Apple Books, B&N, everywhere)
Program: thinkinglabs.academy
Substack: smalaxos.substack.com
Just read all three. The identity layer article hit the hardest. The distinction between an agent that knows your preferences and one that knows who you are — that's not a UX decision. That's the whole architecture.
My version of that is cruder but load-bearing. Claude has a project with 100+ files — my book manuscript, curriculum scripts, assessment data, workbook builds, business strategy docs. When I start a session, it doesn't ask who I am. It knows my vocabulary, my credentials, my student's assessment profile, my business partner's role, my TBI limitations, where I keep files. That's not prompt engineering. That's the identity layer you're describing, built through accumulation instead of architecture.
Your error registry with 3,700+ occurrences and the lesson graduation system — that's what I don't have. My version is me telling Claude "we fixed this before, search the chats" and hoping the right thread surfaces. Your system makes the learning structural. Mine depends on my memory, which is the one thing I can't trust.
The 13-point health check that monitors whether the monitoring systems are working — that's the part I want to steal. My assessment app had a database bug eating results for 23 days and nobody caught it. A meta-system check would have caught it in 24 hours.
Thank you for picking up the book. It's not about AI agents — it's about what happens to human cognition when tools compress the interval where learning occurs. But after reading your architecture, I think you'll see the same pattern from the other side: you're building systems that preserve the human in the loop. I'm teaching students to be the human worth keeping in the loop.
Same problem. Different altitude.
The 13-launcher Llama stack writeup is on the list. Fair trade for what you just shared.
Thanks for sharing this, genuinely. "Survival-built architecture" is the exact phrase. I built most of mine before I had a name for it too, then later read the ADHD systems literature and realized I had been rediscovering the wheel.
Your split is very close to mine. Claude does judgment, local stack does volume. Same load-bearing logic, different hardware.
On the two things you asked:
Night-shift handoff, mechanics here: https://thoughts.jock.pl/p/building-ai-agent-night-shifts-ep1. The original "why" is in an earlier post: https://thoughts.jock.pl/p/my-ai-agent-works-night-shifts-builds.
Memory layer: https://thoughts.jock.pl/p/wiz-ai-agent-self-improvement-architecture. That one is about how the agent keeps knowing who I am across sessions, not just what I want.
Would love to trade notes properly. I will pick up the book. And if you ever write up the 13-launcher Llama stack and the morning-brief pipeline, send a link. Feels like the same problem solved from a different angle.