Dive52
aiclaudediscordautomationobsidiantypescriptproductivitymulti-agent
April 2026 Active

Dive52

A weekly deep-dive learning sprint system — set a topic, drop articles and podcast clips into Discord, react to what resonates, and let Claude synthesize it all into a structured essay that lands directly in your Obsidian vault.

↗ View Live

Built with

Claude
Claude
Obsidian
Obsidian
TypeScript
TypeScript

The Problem

I read 30 articles in a week, flag them, and by Sunday remember approximately none of it. The insight doesn’t stick. The throughput is wasted.

The gap: there’s no forcing function between consuming and integrating knowledge. Reading and understanding are not the same thing. I needed a system that made me take a position on what I was learning — and turned that position into something durable.

Not open-ended “learn more about AI” but “this week: AI data center buildout. Go.” The time-box matters as much as the synthesis. Constraints work.

What Dive52 Is

A 52-week learning curriculum machine. Each week has one topic. Everything you encounter about that topic gets routed through a Discord bot. By Friday, Claude has synthesized your flagged content into a structured essay — stored in Obsidian and pushed to GitHub.

52 dives a year. Every domain that matters.

The Architecture

Input Layer

  • Article URLs dropped directly into Discord — the bot fetches full text, generates a structured card: key ideas, surprising claims, connections to the week’s topic
  • Snipd podcast clips auto-processed from vault/inbox/snipd/ — the bot transcribes, extracts, and cards them
  • Voice notes via Whisper — speak a reaction and it gets processed

Intelligence Layer

  • Claude generates each card with a consistent schema: title, source, key ideas, synthesis potential, relevance score
  • The watcher (chokidar) monitors the vault inbox and triggers card generation on new files
  • Cards live in vault/cards/ with frontmatter: topic_week, reactions, flagged

Curation Layer

  • Discord reactions drive the synthesis:
    • ❤️ = Important (included in synthesis)
    • ⭐ = Definitely synthesize this
    • 🔖 = Read later (resurfaces late week)
    • 👎 = Not relevant (skip)
  • /ask lets you interrogate your sprint mid-week — Claude answers from your actual cards

Output Layer

  • /synthesize gathers all flagged cards + voice reaction notes → Claude generates a 1,200–1,800 word structured essay
  • /publish commits the essay to vault/syntheses/ and pushes to GitHub
  • The synthesis lands in your Obsidian Notebook vault, ready for linking and review

Commands

CommandWhat it does
/topic name:...Set the week’s learning focus
/feed [count:5]Post next N unread cards to channel
/ask question:...Query Claude on this week’s content
/synthesizeGenerate the weekly essay
/publishSave synthesis to Obsidian + GitHub
/statusCards processed, flagged, daily API spend
/pause / /resumeKill switch for all automation

Tech Stack

  • Discord.js — bot interface and slash commands
  • Claude API (Anthropic) — card generation, synthesis, conversational Q&A
  • OpenAI Whisper — voice note transcription
  • chokidar — file system watcher for inbox automation
  • simple-git — vault commits and GitHub sync
  • gray-matter — YAML frontmatter parsing for vault files
  • Fly.io — always-on deployment (~$3/month)

Status

512 articles fetched, 215+ cards generated, running on Fly.io (~$3/month). Active and used weekly.

Up next

All Roads to Kill Bill
All Roads to Kill Bill
An interactive 3D influence network mapping every film that fed into Kill Bill — and every film it spawned. 200+ nodes, live force simulation. Built with React, Three.js, and TMDB.