KNOWLEDGE & WORKFLOWS

Obsidian, Claude and Jev: what an AI second brain actually needs

Understand the roles, build a small pilot and check the evidence before adding continuous automation.

By Superpowers by Waymaker · Published October 23, 2026 · 5 minute read

What is a second brain?

It is a collection of notes and sources organized so you can use them again. An AI-assisted version can help find relevant information and turn it into a draft. Its value depends on whether you can trace the answer back to reliable material.

Karpathy’s LLM Wiki proposes a persistent set of linked knowledge pages maintained from source documents. The viral post adds a particular tool combination. We found documentation for the tools, but have not reproduced the poster’s complete workflow or verified its performance claims.

Give each part a specific job

Obsidian stores Markdown notes in a local folder called a vault. Claude Opus 5.5 is an Anthropic model that can perform knowledge work. Jev, from TypeSafe, returns structured decisions that software can use. These are separate roles, and they need a configured workflow to work together.

One possible design is: find candidate notes, rank their relevance, pass selected evidence to a writing assistant, then review the proposed answer and edits. That design is an editorial example. It is not a claim that these products connect automatically or that Karpathy endorses this combination.

Jev does not have to read every note first

The separate, public Jev Graph Search project makes a local shortlist and sends selected passages to Jev for ranking. It retains source references. Its documentation also says ranking cannot recover a missing candidate or prove that the evidence is sufficient. This is one documented implementation, not verification of the poster’s setup.

That distinction matters for cost and privacy. Inspect exactly what your chosen integration sends, where it sends it and what it writes back. Local storage and local AI processing are different things. Begin with a copied, non-sensitive sample vault.

Build a ten-note pilot

Choose one project and collect ten useful notes: a brief, decisions, references and open questions. Add a date and source to each. Keep the originals intact. Ask your assistant to create a short index and a proposed summary with links, using whatever file access your approved setup supports.

Start with a manual run. You do not need a relevance-ranking model to find out whether ten organized notes help you. Add retrieval tooling when your library and test questions show a need. Each additional component should solve an observed problem.

Use a prompt that makes uncertainty visible

Try the following instruction after giving the assistant access to your permitted sample:

Answer this project question using only the supplied notes. Cite the filename and supporting passage for every factual conclusion. Separate facts, suggestions and unknowns. If the notes conflict, show both versions. Propose a summary update separately. Do not change original notes.

Test the answer and the proposed edits

Prepare five questions with known answers. Include one whose answer is absent and one with conflicting dates. Check source accuracy, unsupported claims, missed evidence and review time. Inspect a before-and-after comparison of every proposed file edit. Keep a backup before allowing writes.

For a recurring run, add a schedule, a spending limit, a run log and a failure notification. Decide which changes can be saved and which require approval. “Thinks with you 24/7” should describe a configured and monitored process, not an assumed property of a notes app.

Your first success is modest and useful: finding the right note, producing a grounded answer and saving a reviewed improvement. Explore a practical workflow playbook before expanding the stack.

Common questions

Is this Karpathy’s official tool stack?

Karpathy published the LLM Wiki pattern. The specific Obsidian, Opus 5.5 and Jev combination in the supplied social post should not be treated as his endorsement.

Is Jev required?

No. Begin with a small manual knowledge workflow. Consider an additional retrieval or ranking component if your tests show a problem it can solve.

Is everything kept on my computer?

Obsidian stores notes locally, but connected AI services may receive selected content. Check the actual integration and provider settings.

Explore the tools

Claude ↗

Sources and review notes

Product information checked against publisher documentation on October 23, 2026. Plans and features can change. The workflows and examples are editorial guidance; this article does not report hands-on product test results.

Keep learning

← All journal articles