GPT-6 Astra inside Obsidian was used to build a knowledge base inspired by Karpathy's LLM Wiki that stores winning ad structures, sales spines, character sheets and timed video sequences for a reusable long-form ad agent.

karpathy drops a wiki mario builds an ad agency in obsidian by dinner

that's a lot of nodes for what's ultimately a copywriting cheat sheet with a nicer interface

Pretty graph aside, the useful part is turning one-off ad prompts into a library you can actually pull from next week instead of starting from zero again.

honestly feels like overkill for ad copy, but i get the appeal

that graph view looks so satisfying

That’s insane

This is how you build actual moats with AI, moving past ephemeral prompt sessions to persistent, interconnected knowledge graphs turns ad generation into a deterministic assembly line.

Now that’s next-level productivity!

Reusable agent memory changes the workload from one-off prompting to continuous retrieval and execution. That is where @fluence_project fits: a verifiable backend for running agent tasks without anchoring the stack to one cloud

Every winning spine stops converting eventually and the agent will keep citing stale notes like gospel. I'd want a kill switch for old structures. How does it flag a note that went cold?

The interesting leap is turning context into a reusable production asset—not just a one-off generation. Versioning those structures and measuring which sequences actually work feels like the hard, valuable layer.

worth checking whether Astra's outputs stay retrievable as separate notes or just live in chat history. if the wiki isn't backed by actual linked files in the vault, your sales spines won't survive an Obsidian reindex or plugin update.

the agent can reuse proven structures while preserving timing, characters, and sales logic, turning ad creation into a repeatable system rather than one-off prompting.

that's the real unlock: collapsing the time between idea and shipped product isn't just about ai or copying ui, it's about collapsing the knowledge gap between disciplines.

External structure beats baking it into weights when the offer or tone shifts mid-campaign.

This is where AI gets really interesting. The real advantage isn’t the model — it’s the knowledge base you build around it. Once your best ideas, structures and workflows become reusable context, every new project starts with your previous wins. That compounds fast. 🧠🤖 Follow for more practical AI systems.

I’m more interested in whether the agent queries the notes directly. If the winning ad spines only show up as pretty graphs, we’re just decorating the workflow.