The Obsidian heads were right.

Tech/product overview

• The video proposes an “agent-powered notes → wiki + visualization” workflow (targeted at “Obsidian heads”).
• Core idea: keep capturing messy notes as Markdown files, then use AI agents (“skills”) to:
  1. Enrich notes with metadata (tags, source URLs, backlinks/links).
  2. Generate/update wikis (index pages + concept/person/source maps).
  3. Create HTML visualizations (charts/graph views) from the same note corpus.
• Emphasis: systems should update automatically “while you sleep,” so the outputs stay fresh without manual upkeep.

---

Phase 0: Capture raw notes (Markdown-first)

• Use a Markdown “raw folder” as the intake directory.
• Notes can come from multiple sources:

Voice transcription (suggested as most efficient)

• Tools recommended:
  • Handy (free/open source): local/on-device transcription using an on-device model (mentioned: “Nvidia Parakeet” at recording time).
  • VoiceInk (open source): preferred because it generates paragraph breaks.
    • Costs about one-time ~$20 for updates.
    • Supports custom models and claims on-device behavior.
    • Also mentions using context from the screen to improve transcription.

Web clipping

• Obsidian Web Clipper:
  • Converts webpages into Markdown with front-matter properties (e.g., author, tags).
  • Copies the source URL.
• Also works with YouTube transcripts:
  • Imports transcripts with timestamps.

---

Phase 1: Enrich notes with metadata (Enrich Note skill)

• Introduces an agent skill called “Enrich Note.”
• Enrichment goals:
  • Add tags from an existing organization system (stored as Markdown).
  • Add source details (e.g., find the podcast URL via web search).
  • Add backlinks/related links using Wikilink format.
• Important constraint:
  • The agent is instructed to be “reluctant” to invent new tags, preventing tag sprawl.

Implementation approach

• Because notes are already Markdown, enrichment can run by reading/editing files directly—no special connectors/MCP servers required (it “routes to whatever that directory is”).
• For efficiency, recommends a low-reasoning / token-efficient model for these enrichment skills.

---

Phase 2: Generate and maintain wikis automatically

Wiki generation concept

• Ask an agent to generate:
  • A wiki index page
  • A map of:
    • people
    • concepts
    • sources/material
  • Clickable pages for deeper exploration.
• Template reference:
  • Mentions copying a wiki prompt/template from a GitHub gist by Andrej Karpathy to produce topic-specific wikis.

Example wiki outputs mentioned

• “Latest in AI” index/wikis
• Pages for:
  • transcription tools
  • Claude Code
  • side projects
  • voice transcription tooling
  • AI handoff strategy
• Unexpected entries (example):
  • Adam Neely due to AI influence in music
• Organization/company entities relevant to the topic.

Automation/refresh behavior

• Instead of manually generating wikis, the system should:
  1. Run an agent to enrich/tag raw notes
  2. Run another agent to refresh relevant wikis
• Result: when you read a wiki later, it’s fresh.

---

Cloud automation & syncing (keep agents “running”)

Recommended architecture

• Run agents in a cloud sandbox/container.
• Use file sync to copy notes between local and cloud, execute skills, then sync back.

Sync option (preferred)

• Obsidian Sync (paid): about $4/month.
• Uses OB / “Headless Obsidian” CLI with a single sync command to pull/fetch Markdown into a sandbox.

Alternative option

• Use git to clone/push notes (free, but more hassle).

Example cloud setup

• Using Warp team’s Oz cloud platform with Docker:
  • Container includes Obsidian CLI (+ JavaScript dependency for the CLI)
  • Agent flow: clone → sync down → run wiki/enrichment skills → sync up

Scheduling + idempotency

• Use a loop that marks notes as already processed using front-matter:
  • If “already enriched,” the agent skips previously processed notes.
• Cadence example:
  • Daily (or weekly to save tokens)

Tooling visibility

• Oz provides a browser preview showing edits, enriched resources, and job cadence.

---

Visualization generation (HTML outputs)

• Beyond text/link navigation, the speaker demonstrates HTML visualizations generated from notes:

Burndown chart

• Like GitHub commits:
  • Tracks daily streak/notes activity
  • Displayed via an in-app HTML viewer

Graph/constellation view

• Inspired by Obsidian graph view, themed as “outer space.”
• Features:
  • Shows clusters/hotspots (e.g., AI/engineering, Claude Code-related nodes)
  • Click through to nodes and related notes
• Customizable:
  • Ask agents to change the visualization; use creative layouts/themes.

Encouragement

• Use agents + coding agents to generate HTML files (e.g., “generate an HTML file for me”) from the Markdown corpus.

---

Product mentioned

• Hubble: a note-taking/editor app created by the speaker.
  • Provides an Apple Notes-like UI while writing content to Markdown.
  • Described as free/open source and cross-platform.
  • Used for voice-note capture in the workflow.

---

Main speakers / sources (as stated or implied)

• Speaker: the narrator/creator of the workflow (also builder of Hubble; mentions “I built it in my garage”).
• Template/source referenced: Andrej Karpathy (GitHub gist used as a wiki-generation template).
• Team/source referenced: Claude team (mentioned in relation to HTML visualization possibilities).
• Tool authors/platforms referenced:
  • Obsidian (web clipper, sync, Obsidian CLI/Headless Obsidian)
  • Warp (Oz) (cloud runner)
  • Transcription apps: Handy and VoiceInk

---