The Night School: How My AI Reads 54 Playbooks Overnight and Gets Smarter
What It Is
I’ve got a stack of ops books and PDFs I bought years ago and barely opened. So I built a thing that reads them for me while I sleep. Every night at 2 AM it kicks off, pulls the 54 playbooks off my NAS, chunks them up, and feeds them through a local Ollama model. By morning I’ve got a fresh vector store and some rough notes I can search. It won’t replace actually reading them. But it beats letting them collect dust.
What’s Actually Running
Here’s the short version: n8n orchestrates the job, Python handles the chunking, Ollama does the embeddings and summarization, and the output lands in a local vector database on the same box. Nothing leaves my network.
I went with Ollama because I don’t want my books passing through someone else’s API. Even with a “we won’t train on this” policy, I’d rather not bet on it. Local inference is slower, sure. But the whole thing kicks off at 2 AM when I’m not using the GPU for anything else, so the speed hit doesn’t matter to me.
How the Night Goes
Six hours start to finish. The embedding pass eats most of that. Summarization moves quick because I’m running Mistral 7B and asking for one paragraph per chapter, nothing fancy.
I log what ran and what didn’t. Maybe one in ten nights something chokes. Usually a PDF with weird encoding or a book that’s mostly scanned images. Those get pushed to the weekend.
That’s the tradeoff. Faster nights mean I’m cleaning up a failed queue the next morning. I’d rather lose an hour Sunday than have this drag into a 14-hour run that bleeds into my work day.
Does It Actually Make the AI Smarter?
Not really. Not the way you or I get smarter after finishing a good book. The model just ends up with a bigger haystack to pull from when I ask it stuff.
Take Kubernetes networking. I ping the AI about a weird CNI issue and it surfaces stuff from the books I’ve fed in, mixed with my own notes. That’s faster than me fumbling through 54 PDFs at 2am looking for the one with the decent diagram. I still know more than it does on most topics. The win is that it remembers more of what I’ve already learned.
The Tradeoffs
This isn’t free. I mean that in a few ways at once. There’s the hardware cost, the electricity, and the fact that the box is useless for anything else while a 54-doc run is grinding through. My machine gets loud enough that I can hear it from the next room when it’s working hard. Not exactly a fire-and-forget setup.
The Concept
Every night at 8 PM, my AI agent kicks off its homework. It opens a playbook, which is just a short guide I wrote on some AI tool, automation pattern, or business model. It pulls the takeaways and shoves them into semantic memory. By morning I can ask it anything and it cites whatever’s relevant, no file-fishing needed. I built this because I got sick of feeding the same context over and over.
The queue sits at 54 playbooks right now. 31,884 words total.
What Is a Playbook?
A playbook isn’t a blog post. It’s not a tutorial either. It’s a tighter format I built for this site, and here’s how I define it:
- Practical – “Here’s exactly how to do X”
- Tested – Based on actual experiments, not theory
- Concise – 500-2,000 words, no fluff
- Categorized – AI/ML, automation, business, income, creative
Topics run from “Solana Trading Basics” to “n8n Workflow Automation” to “AI Creative Writing Services.” Each one is a skill or business model I’ve researched and boiled down. The n8n ones come straight from workflows I actually run on this site. The trading and creative writing ones lean on vendor docs, community threads, and a fair bit of trial and error, and I’ll flag which is which in each post so you know the difference.
The Storage System
I run LanceDB locally with Ollama’s nomic-embed-text model for embeddings. Nothing exotic, nothing cloud-based.
Why local?
- No API costs for storage or retrieval
- No rate limits to fight
- Data never leaves my machine
- ~200ms query time on semantic search
Yeah, you manage it yourself. That’s the deal.
How it works:
- Read the playbook text
- Generate a 768-dim embedding through Ollama
- Store it in LanceDB with metadata I want (topic, category, date)
- Query by semantic similarity when I need it
Setup took about 20 minutes. The only annoying part was figuring out the metadata schema, and that’s a one-time tax.
Real Queries
I ran a handful of searches this week against my own index. Real questions, real output.
Query: “How do I make money with AI agents?”
Result: Income streams playbook (score: 0.89), AI agency income (score: 0.85), Freelancer platforms (score: 0.71)
Query: “What GPU do I need for local LLMs?”
Result: AI/ML playbook (score: 0.92), Nova trading bot (score: 0.67), ComfyUI notes (score: 0.61)
Query: “How to automate social media posting?”
Result: Mixpost playbook (score: 0.88), Postiz playbook (score: 0.84), RSS reader (score: 0.72)
It’s not keyword matching. The GPU query pulls the AI/ML playbook even though that file never says “GPU” or “LLM” in the title. That’s the semantic layer earning its keep.
Not perfect though. The Nova trading bot showing up on a GPU query tells me my embeddings cluster too tight around anything compute-related. Still tweaking the chunk size. Probably going to throw some negative examples at it to push that one down the rankings.
That’s the tradeoff with semantic search. You get the “GPU” to “AI/ML” connection for free, but you also pick up weird neighbors that share a vibe instead of a topic. Annoying, but fixable.
The Nightly Routine
Every night at 8 PM, a cron job fires on a small VM. Here’s the whole loop:
1. Pull a playbook from the queue (docs/night-school/queue/)
2. Read it and extract 5 takeaways
3. Save the summary to LanceDB with a vector embedding
4. Post a note in Discord — “Tonight I learned about [topic]”
Run that for 54 nights and you’ve got 54 chunks the agent can reason over, not just retrieve. Retrieval hands you a link. Reasoning hands you an answer. That gap is bigger than it sounds, and I didn’t expect to care about it until I saw the difference in practice.
It’s not magic. Garbage playbooks in, garbage summaries out. You still have to curate the queue, or you’ll spend a week chasing weird answers before realizing the source material was junk to start with.
What Changed After 54 Playbooks
The AI doesn’t pull from generic training data anymore. It pulls from playbook comparisons I actually wrote. “Use n8n instead of Zapier for self-hosted workflows” came from one of mine, not from some scraped web page.
Drafting posts got faster too. When I started one last week, the AI pulled from the n8n, automation-comparison, and openclaw-subagents playbooks at the same time. The first drafts come out more detailed than they used to, though I still edit the hell out of them before anything goes live.
The business angle is where it surprised me. Stack the income-streams playbook with the AI agency one and the freelancer platforms one, and you’ve got a coherent strategy for selling AI automation services. On their own they’re useful. Together they read like a plan.
One thing I won’t pretend: it still gets things wrong. Pulls from the right playbook, then misquotes a pricing figure or invents a feature that doesn’t exist. I catch it in editing. If I wasn’t paying attention it would ship. That’s the part nobody mentions when they brag about AI workflows.
The Viewer
All 54 playbooks live at http://192.168.68.51/night-school-viewer.html on my Synology NAS. Open it from anything on the network. Laptop, phone, doesn’t matter.
What it does:
- Sidebar navigation by category
- Markdown rendering
- Search (planned)
- Edit links that jump straight to the local files
It’s not pretty. I spent maybe ten minutes on the CSS before I gave up. Styling is something I tolerate, not enjoy. But the playbooks load fast, and when I edit a file I refresh and see the change in seconds. That’s what I actually care about.
The playbooks are the product. The viewer is just a way to read them without dropping into a terminal.
Building Your Own Night School
You don’t need 54 playbooks to start. Here’s the minimum version that actually works.
1. Pick 5 topics you want to learn
2. Write 500 words each on what you learned, what worked, what didn’t
3. Store in any database — a text file with headings is fine
4. Query before decisions with prompts like “What do I know about X?”
No fancy infrastructure needed. The point is to start writing, not architect the perfect system.
For the vector memory version, this is the stack I run at home:
- LanceDB (free, local)
- Ollama + nomic-embed-text (free, local)
- 100 lines of Python
Total cost: $0. Setup took me about 2 hours, and yeah, half of that was chasing a dependency conflict I should’ve caught earlier. That’s the tradeoff nobody talks about with “local AI stacks.” It’s free, but it eats your time if you’re not careful.
What’s Next
The agent works. Now I want it sharper. Here’s where I’m headed:
1. Inter-playbook linking: so the agent flags “this playbook references that one” instead of leaving those connections buried in old files.
2. Spaced repetition: re-surface old playbooks when a current project lands in the same territory. Anything older than a few months is dead weight, and I’ve already got plenty of notes I’ve forgotten I wrote.
3. Collaborative learning: run multiple agents in parallel, each reading different playbooks, then have them swap notes on what they found.
I’ll write it up as it comes together.
The Realization
I write playbooks for my AI, but honestly the playbooks are for me first. Writing them out forces me to notice what I actually learned, not just what I thought I learned. Then the model reads them overnight and we both end up sharper the next morning.
It’s not a course. It’s a working back-and-forth between me and the model, testing what holds up and what doesn’t.
Want the Night School setup? I’m packaging the vector memory system plus 10 starter playbooks as a downloadable ($35). LanceDB config, Ollama integration, and a viewer included so you can poke around the knowledge base yourself.
The tradeoff is real though. You have to write the playbooks yourself. Feed it garbage and you’ll get garbage back. But do the work upfront and it pays you back every night while you sleep.
—
This is how I run my business now. One playbook at a time, one night at a time, watching a shared knowledge base get a little smarter every morning.