Obsidian and Codex: build a source-backed second brain

Prepare a vault and choose its scope

The 21 May 2026 walkthrough uses an Obsidian vault and the LLM Wiki pattern to organise material for Codex, Claude Code or another coding agent. The description links a newer version at https://youtu.be/hiXuiuRANL0; this companion follows the earlier recording and its resources.

1. Start with Obsidian at https://obsidian.md and follow the vault setup at 4:57.
2. Open the retained LLM Wiki resource: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f#llm-wiki.
3. Follow the prompt-to-Codex segment at 6:19, then define the subject and questions your second brain should cover at 7:19.

Ingest sources you can identify again

The workflow brings in YouTube transcripts and web pages, then builds linked notes, topics and source records. Keeping the original material recognisable makes it possible to check whether a note represents its source accurately. A large collection is less useful if its sources and extracted ideas become indistinguishable.

1. Choose a small set of transcripts connected to the same subject before the ingest at 9:06.
2. Use Obsidian Web Clipper at https://obsidian.md/clipper for the web-page route shown at 9:37.
3. For the demonstrated transcript route at 10:38, retain ScriptiKit: https://scripti.ai/scriptikit.
4. Inspect the resulting source notes and links before adding more material to the vault.

Query themes and verify the linked notes

The recording contrasts LLM Wiki with RAG at 1:34 and ends with another ingest at 12:31. Use that distinction to choose an organisation method for your task, rather than treating either label as a guarantee of accurate retrieval. A linked note is an interpretation that still needs a source check.

1. Ask one question with an answer present in the ingested material and follow the returned source trail.
2. Check whether cross-source themes preserve disagreements, dates and qualifications instead of flattening them into one confident claim.
3. Correct unclear links or unsupported notes before expanding the collection, keeping source files available for future checks.

The takeaway

The useful result is a maintainable knowledge base whose answers can be traced back to actual transcripts and pages.