🧠 Second Brain — Stretch Goals
Finished the main Second Brain tutorial early? Pick any of these.
1 · Linting pass
Ask the agent to do a health check on the wiki you just built.
Paste into opencode:
You are auditing the wiki/ directory. Produce a short markdown report at
wiki/_lint-report.md listing:
1. Files missing required frontmatter fields (title, tags, sources).
2. Wiki-links that point to non-existent files.
3. Paper summaries that are noticeably shorter or longer than average.
4. Concept articles that link to fewer than 2 papers (probably too thin).
5. Themes that appear across 3+ papers but have no concept article yet.
6. For each issue, a one-line fix suggestion.
Do not change any files yet. Just write the report.
Read the report, agree or disagree with each finding, then ask the agent to apply the fixes you approve.
2 · Claude-Obsidian plugin
The claude-obsidian plugin lets Claude chat directly inside Obsidian, alongside opencode. Use opencode for long-running compile/edit jobs, and the plugin for quick conversational lookups without leaving the vault.
- In Obsidian:
Settings → Community plugins → Browse, search “Claude”. - Install and enable the plugin.
- Paste an Anthropic API key (different from your OpenRouter key — ask Jay for one if you don’t have your own).
- Open the plugin’s side panel and ask: “Summarise the concept article on [topic] in two sentences.”
The plugin reads the current note as context. Handy for “what did I write about this again?” questions.
3 · Matplotlib figure from the wiki
Have the agent write a small Python script that reads wiki metadata and renders a figure.
Write scripts/citation_timeline.py that:
1. Walks wiki/papers/*.md and parses the YAML frontmatter.
2. Extracts the publication year from each paper.
3. Renders a histogram of papers by year using matplotlib.
4. Saves the PNG to wiki/figures/citation_timeline.png.
Then run it and show me the result.
Other figure ideas:
- Concept co-occurrence matrix — which concept articles share paper references?
- Tag cloud — frequency of tags across
wiki/papers/*.md. - Backlink graph — Obsidian has a native graph view; you can also export via the agent.
Drop the figure back into Obsidian and [[link]] it from INDEX.md.
4 · Marp slide deck
Turn your cross-paper Q&A answer into a slide deck you could present.
Read wiki/questions/<your-answer>.md. Convert it into a Marp slide deck
at wiki/slides/<your-answer>.md with:
- Title slide
- One slide per main point
- Quote slides for direct evidence
- A "further questions" slide
- Speaker notes under each slide
Use Marp's front-matter format so it renders in the Marp for Obsidian
plugin or Marp CLI.
Install Marp for Obsidian or Marp CLI to preview and export to PDF / HTML.
5 · Hook up a search engine
Karpathy’s original post mentions building a naive search engine over the wiki as a tool. This is a 30-min job with an agent.
Build scripts/search.py that:
1. Walks wiki/**.md and builds a simple TF-IDF index in memory.
2. Exposes a CLI: `python scripts/search.py "query terms"` prints
the top 5 matching files with a 1-line snippet.
3. Also writes an HTML interface at wiki/_search.html that I can
open in a browser.
Then test it with 3 example queries.
Once this exists, you can give the agent CLI access to your search — “use scripts/search.py to find all papers mentioning X, then write a concept article on them” — and queries become much cheaper than re-reading every file.
6 · Incremental growth loop
Set up the routine so the vault grows without you managing it.
Create scripts/add-paper.sh:
#!/usr/bin/env bash
set -euo pipefail
paper="$1"
cp "$paper" raw/
echo "Added $paper to raw/. Now run the compile prompt in opencode."
Or just tell the agent: “Every time I drop a new PDF in raw/, do the following without asking: write its summary, update the index, and update any concept article that touches it. Wait for my approval before merging concept changes.”