| name | vault-compare |
| description | Compare the 4 knowledge management approaches and recommend the best fit for the user. Use when the user asks 'which approach should I use?', 'compare the approaches', or is unsure how to start. |
| allowed-tools | Read Glob Grep |
Vault Compare
Help the user choose between the 4 Obsidian + AI knowledge management approaches.
Workflow
Step 1: Understand the user's situation
Ask about:
- Content type: Books? Articles? YouTube videos? Meeting notes? Code docs? Podcasts?
- Volume: How much content do they have? How fast does it grow?
- Primary use: Retrieval? Learning? Token reduction? Search?
- Technical comfort: CLI tools? Obsidian experience? Zettelkasten knowledge?
- Time budget: 5 minutes? A weekend? Ongoing commitment?
- Who consumes the output: Human browsing? AI agent? Both?
Step 2: Present the comparison
Use this scoreboard:
| Dimension | Second Brain | Personal Search | AI Coding Vault | Karpathy LLM Wiki |
|---|
| Technical Complexity | 6/10 | 4/10 | 6/10 | 2/10 |
| ROI | 7/10 setup, 5/10 long-term | 8/10 | 7/10 (conditional) | 9/10 setup, 7/10 long-term |
| Time to Value | Days | 1 week | 6-12 weeks | Minutes to hours |
| Primary Consumer | Human (thinking) | Human (retrieval) | AI agent (navigation) | Both human + AI |
| Maintenance | High (curation) | Low (~1hr/month) | Medium (session-end) | Low (periodic lint) |
| Failure Mode | Vault bloat + false confidence | Blind spots | Stale docs mislead agent | Index decay |
| Human Curation | AI extracts, human reviews | Human curates sources | Human defines structure | AI does everything |
| Infrastructure | CLI + Obsidian | Scripts + commands | SKILL.md + conventions | Just markdown |
Step 3: Make a recommendation
Choose Karpathy's LLM Wiki if:
- You want to start in 5 minutes, not 5 days
- You have raw data piling up
- You want AI to handle all organization
- Your scale is personal (hundreds of pages)
Choose Personal Search Engine if:
- You already follow 5-15 trusted creators
- You want ROI this week
- You've tried Zettelkasten and abandoned it
Choose Second Brain Building if:
- You consume 2-4+ books/month
- You want AI that cites YOUR reading
- Your domain changes slowly
Choose AI Coding Vault if:
- You use Claude Code daily on a 3+ month project
- You've copy-pasted architecture explanations 5+ times
- Token costs are a real concern
Step 4: Present the layering strategy
If the user has multiple needs, recommend the layered approach:
| Vault | Purpose | When to Start |
|---|
| Karpathy Wiki | Raw data → organized knowledge | First (lowest effort) |
| Mentor Vault (PSE) | Search trusted sources | When you have 5+ trusted sources |
| Project Vault (AI Coding) | Per-project agent memory | When a project hits 3+ months |
| Library Vault (Second Brain) | Book/course knowledge | When you have a book backlog |
Start with one. Add others as needs grow.
Step 5: Karpathy LLM Wiki vs Semantic Search RAG
If the user asks about RAG, present this comparison:
| Dimension | LLM Wiki | Semantic Search RAG |
|---|
| How it finds info | Reads indexes, follows links | Vector similarity search |
| Understanding | Deep relationships via links | "These chunks seem similar" |
| Infrastructure | Just markdown files | Embedding model + vector DB + chunking pipeline |
| Cost | Free (token costs only) | Ongoing compute + storage |
| Maintenance | Run a lint, add articles | Re-embed when things change |
| Sweet spot | Hundreds of pages | Millions of documents |
Scale threshold: hundreds of pages with good indexes = wiki. Millions of documents = traditional RAG.
