# Using LLM Wikis to Prepare Courses In September 2026, [[Matt Pocock]] shared the workflow he uses to plan courses with AI agents. It builds on [[Andrej Karpathy]]'s [[LLM Wiki]] pattern. ## Why he needed a workflow A few days earlier, he pointed out that knowledge work is MUCH harder to automate with agents than code. Software development is easy mode for agents: - Tons of high-quality automated feedback loops (types, tests) - Well-organized documentation, searchable on the web - Simple workspaces Course planning has none of that. No compiler tells you a lesson is confusing. So you have to build the structure and the feedback loops yourself. ## The workflow 1. **One LLM wiki per deliverable**. A deliverable is a sizeable piece of work that takes more than a day: a course, a talk, a video. Each one gets its own Karpathy-style wiki 2. **Split the deliverable into sections** that can be worked on in independent threads during the "day shift". Multiple agent threads run in parallel, all reading from and writing to the wiki, which acts as their shared state 3. **Input through braindumps**. He dictates huge chunks of text, and the model integrates them into the wiki. The model produces HTML reports, and he reacts to those reports by dictating again 4. **A linting skill for the "night shift"**. It runs over the entire deliverable with LOTS of subagents, looking for every possible weakness His summary: he speaks what he wants into existence, the model captures everything he says, and the deliverable slowly comes to life. He calls it a work in progress. ## My thoughts This is very close to how I think about AI-assisted creation, and there's a lot to steal here: - **The wiki is the shared memory.** Parallel agents don't need to talk to each other. They coordinate through the knowledge base. That's the same idea as a [[Personal Knowledge Management (PKM)|PKM]] system acting as a single source of truth for everything else - **Scoping a wiki per deliverable** keeps context focused. A giant, everything-wiki would pollute every thread with irrelevant information - **Dictation removes the input bottleneck.** Speaking is much faster than typing, and braindumps capture half-baked ideas before they vanish. It's the same reason I rely on voice capture with [[Knowii Voice AI]] - **Linting is the missing feedback loop.** Code has tests. Courses don't. A generous, adversarial review pass at night is the closest thing to a test suite for knowledge work - **Day shift / night shift** is a smart way to split interactive work (where you're needed) from long-running autonomous work (where you're not) One thing to watch: Steven Thompson's critique of LLM wikis (compilation vs curation) applies here too. The model can capture and structure everything you say, but deciding what actually matters for learners stays your job. I could apply this almost directly to my own courses, like the [[Obsidian Starter Course]]: one wiki per course, one thread per module, voice braindumps as input, and a nightly review with the `osk-wiki-lint` approach. ## References - Matt Pocock's thread: https://x.com/mattpocockuk/status/2097638166232457451 ## Related - [[LLM Wiki]] - [[Andrej Karpathy]] - [[Matt Pocock]] - [[AI Wiki PKM]] - [[AI Subagents]] - [[Claude Code Skills]] - [[Knowii Voice AI]] - [[Codebase to Course]] - [[Obsidian Starter Course]] - [[Knowledge Workers]]