Build a Self-Improving Claude Knowledge Base with ONE Prompt

Main Idea

The video presents a “self-improving Claude knowledge base / second brain” that can be created with a single prompt. It focuses on building it without complex tooling—specifically no Obsidian, no vector database, and no RAG setup—using only Claude plus one well-crafted prompt.

Setup & Tooling (Tutorial-Style)

• Goal: Create a knowledge base that becomes more capable over time (e.g., better on “day 100” than “day 1”).
• Required tools: Claude desktop app (the demo uses the desktop app; the creator personally uses the VS Code extension).
• One-prompt automation: The prompt instructs Claude to:
  • Create the local folder structure
  • Generate all necessary files end-to-end, including configuration and routines
• Cloud Code confirmation: Claude uses Cloud Code to confirm actions and then automatically perform file creation.

Core System Architecture (5 Building Blocks)

The knowledge base is described as five main components, mostly folders + text/Markdown files:

1. raw/ (inputs)

   • A messy dumping ground for:
     • Research files
     • Captured articles/sources
   • Inputs are intentionally unorganized at first; organization happens later.

2. wiki/ (organization layer)

   • The agent organizes raw inputs into searchable linked knowledge.
   • Includes mechanisms such as:
     • an index
     • a log
   • Acts as the navigable map Claude uses to understand what exists and where.

3. outputs/ (generated artifacts)

   • Where the system places generated results such as:
     • PDFs
     • slide packs
     • applications
   • Generated from the combination of wiki/ + raw/ inputs.

4. claw.md (operating manual / system prompt)

   • A Markdown instruction file functioning like an operating manual for the agent.
   • Best practice attributed to Andre Karpathy:
     • claw.md should be maintained by the agent itself
     • This enables self-evolution of instructions over time
   • The prompt reportedly incorporates Karpathy’s “LLM Wiki pattern.”

5. “Dream sequence” routine (self-audit + self-update)

   • A scheduled lint/health-check + maintenance process.
   • Initially described as weekly, later adjusted to daily in the demo.
   • It:
     • Ingests new raw material
     • Runs a lint step to catch issues like contradictions, stale claims, duplicates, or orphans
     • Updates the knowledge base so it stays current

How the “Dream Sequence” Works (Maintenance Behavior)

• The demo shows the routine managed via the Claude app UI under “routines.”
• It can also be manually invoked using natural language.
• Manual run procedure described:
  • Claude compares files in raw/ vs a process registry
  • Then merges/folds new items into wiki/
  • Updates connections so future queries are easier

Knowledge Ingestion Workflow (Examples)

• Deep research ingestion:

  • For the demo topic (“AI and the upcoming agentic industry”), Claude performs research and stores results in raw/.
  • It then updates wiki/ so relevant sources can be retrieved later.

• Manual link saving:

  • The video describes a workflow where the user shares an article link to Claude and asks it to save.
  • Claude saves the link into raw/ as a Markdown file and updates wiki/ accordingly.

• Querying the knowledge base:

  • The creator demonstrates querying “summary notes from Google I/O” that were previously saved.
  • Claude retrieves from both:
    • wiki/ (organized, searchable knowledge)
    • the specific raw/ files containing the original notes

wiki/ Internal Files (What Makes It Navigable)

Within wiki/, three highlighted files:

• index

  • Functions like a table of contents
  • Lists concepts/entities/sources to speed up agent browsing

• log

  • Records:
    • creation timestamps
    • what knowledge was added or removed

• Processed/registry-like markdown file

  • Tracks which raw files have already been ingested
  • Prevents reprocessing the same inputs in later dream sequences

Customization Options (Prompt-Driven Tweaks)

The video claims you can change behavior with additional prompts, including:

• Whether wiki/ auto-updates immediately after ingest
• Whether updates occur only via the dream sequence
• Changing dream sequence frequency (demo changes from weekly to daily)

Promotional / Referenced Resources

• Mentions Robo Nuggets community:
  • Live build sessions
  • Templates referenced as related to what was shown
  • A structured path from “zero to mastery” in agentic AI, including how to get paid

Main Speakers / Sources

• Speaker/creator: The YouTube narrator/host (name not explicitly given in subtitles).
• Primary referenced source: Andrej (Andre) Karpathy (his “LLM Wiki pattern” / viral knowledge base post).
• Tools/platform referenced: Claude (Claude desktop app and Cloud Code routine UI).