Agentic AI for Neuroscience

Second Brain · Build a research wiki with your agent
22 April 2026 · UKDRI · Skene Lab · Imperial
Today

Two hours. Two parts.

30 min · Will Leeney

Supercharging research with LLMs and agents. Talk + Q&A.

90 min · Workshop

You build a working second brain on your laptop. Agent compiles a wiki from 3–5 of your own papers, then answers a cross-paper question.

By the end you leave with a vault that is ready to keep growing after the workshop.

The idea

Stop manipulating code. Start manipulating knowledge.

"A large fraction of my recent token throughput is going less into manipulating code, and more into manipulating knowledge (stored as markdown and images). The latest LLMs are quite good at it." — Andrej Karpathy, LLM Knowledge Bases

The agent is better than you at reading every paper in the folder every time. You stay in the loop by reading, critiquing, and asking questions.

The pattern

Raw → Wiki → Q&A

1. Ingest

You drop PDFs, clipped articles, images into raw/. You never edit raw/.

2. Compile

Agent writes wiki/: per-paper summaries, concept articles grouping themes, backlinks, an INDEX.

3. Query

Agent reads wiki and raw together, answers cross-cutting questions, files the answer back for next time.

Karpathy's vault: ~100 articles, ~400K words. Works without fancy RAG at that scale.

Vault layout

Two folders, one rule

second-brain/ ├── raw/ ← you put things here │ ├── 2024-karpathy.pdf │ ├── 2023-review.pdf │ └── clipped-article.md ├── wiki/ ← agent writes here │ ├── papers/ ← one .md per source │ ├── concepts/ ← themes grouped across papers │ ├── questions/ ← rendered answers │ └── INDEX.md └── AGENT.md ← agent's working instructions

Rule: you own raw/, the agent owns wiki/. You rarely touch wiki/ directly.

The stack

Three tools, one folder

You read in
Obsidian
Viewer. Renders markdown, shows the backlink graph, clips web articles, previews figures.
You talk in
opencode
Editor. Reads raw/, writes wiki/, runs scripts, answers questions.
Powered by
OpenRouter
Provider. Hosts the model. Shared key handed out for today's session.
↑ viewer · editor ↔ agent · model ↓

Both Obsidian and opencode point at the same folder. Edit in one, see it in the other.

The plan

90 minutes · 6 phases

PhaseWhatOutcome
01 · 5 minKickoffGrab the OpenRouter key from Slack. Hear the stack overview.
02 · 20 minInstallObsidian + opencode installed. Key pasted. Empty vault open in both.
03 · 10 minIngest3–5 of your own PDFs in raw/. Optional web clip.
04 · 25 minCompileAgent writes wiki/: summaries, concept articles, INDEX, backlinks.
05 · 20 minQ&AOne cross-paper question answered as markdown or Marp slides.
06 · 10 minSharePost one surprising connection to Slack.
Phase 01 · Kickoff

Your API key is in Slack

Shared OpenRouter key

Posted as a pinned message in #agentic-ai-for-neuroscience-workshop. Copy it now so you can paste in the install phase.

Not in the workshop Slack yet?

Scan the Workshop Slack QR on the landing page: neurogenomics.github.io/agentic-life-sciences-tutorial/workshops/agentic-ai-neuroscience-22-april-2026

Key rotates after the session. Don't commit it to git.

Phase 02 · Install (20 min)

Get the stack running

  1. Obsidian — download from obsidian.md, install, skip the sign-in.
  2. opencode — download from opencode.ai. See the Get OpenCode tutorial on the site for screenshots.
  3. Make a folder: ~/Documents/second-brain/raw/
  4. In Obsidian: Open folder as vault → pick second-brain/.
  5. In opencode: open the same folder as your project.
  6. Add your key: Settings → Providers → OpenRouter → paste.

Get unstuck in Slack · #agentic-ai-for-neuroscience-workshop

Phase 03 · Ingest (10 min)

Drop in your papers

  • 3–5 PDFs of papers you actually want to think about.
  • Review articles + primary research is a good starting mix.
  • Filenames matter: 2024-karpathy-llm-kb.pdf beats s41588-024-01919-z.pdf.
  • Optional: install Obsidian Web Clipper and clip one article directly into raw/.

If you have nothing to hand, use 3 papers from Will's talk. If you're stuck, ask a neighbour.

Phase 04 · Compile (25 min)

One prompt. Let it run.

Read AGENT.md. Then:

1. Scan raw/ and list every source you find.
2. For each source, write wiki/papers/<filename>.md with:
   - frontmatter (title, authors, year, tags, source: raw/<filename>)
   - a 5-sentence summary
   - 3-5 key findings as bullets
   - a "Connections" section with [[wiki-links]]
3. Identify 3-5 themes connecting multiple papers. Write one
   wiki/concepts/<theme>.md per theme, linking back to papers/.
4. Write wiki/INDEX.md listing everything, grouped by tag.

Ask before running if anything is unclear. Otherwise, go.

Full prompt + AGENT.md template are on the tutorial page.

Phase 04 · Compile

Watch Obsidian fill up

In opencode

You see the agent reading PDFs, drafting summaries, proposing concept clusters. Read along — you'll learn what it noticed.

In Obsidian

Files appear live. Click any [[wiki-link]] to jump between papers and concepts. Open the graph view.

If it's slow, tell the agent to do 2 papers now and the rest in a second pass. Iterative is fine.

Phase 05 · Q&A (20 min)

Ask something no single paper answers

Method overlap

"Which papers share upstream methods or datasets? Where do they disagree?"

Repeated limitations

"What are the 3 most-repeated limitations across my review articles?"

Intro references

"If I was writing an introduction to this topic, what 5 references would I cite and why?"

Ask the agent to render the answer as wiki/questions/<slug>.md with supporting quotes and backlinks. Open it in Obsidian. Verify the evidence against the original papers.

Why this works

Three boring superpowers

1. Reads everything

The agent reads every paper in the folder every time. You'd skim. It doesn't.

2. Stays consistent

Frontmatter, backlinks, INDEX. Boring but compounding.

3. Re-compiles

Add a paper, re-run the prompt. Concept articles update themselves.

Scale

From 5 papers to 500

~100
articles in Karpathy's vault
~400K
words total

At this scale, no fancy RAG needed. The agent auto-maintains INDEX + brief summaries and reads related docs as it works.

Phase 06 · Finished early?

Stretch goals

Linting pass

Agent audits your wiki — missing frontmatter, thin concepts, orphan notes.

claude-obsidian

Chat with Claude directly inside Obsidian alongside opencode.

matplotlib

Agent writes a script to plot citation timeline from YAML frontmatter.

Marp slides

Turn your Q&A answer into a slide deck you could present.

Search engine

Vibe-code a TF-IDF CLI the agent can use as a tool.

Growth loop

Script add-paper.sh so new PDFs trigger a compile pass.

Full details: tutorials/10-second-brain-stretch on the workshop site.

Phase 06 · Share (10 min)

What surprised you?

Post one connection the agent found that you didn't expect to the workshop Slack:

  • A citation two papers share that you'd missed.
  • A hidden thematic overlap.
  • A contradicting finding across papers.
  • A concept article that grouped sources you wouldn't have grouped.

The "why did you group these?" reveal is often the best part.

After today

Keep the loop running

  1. Drop a new paper into raw/ → ask the agent to update the wiki.
  2. Have a new question → agent writes the answer into wiki/questions/.
  3. Periodically ask the agent to lint — find inconsistencies, suggest new concepts.

Over weeks, the wiki becomes specific to your research. Every question you answer adds to the next answer's context.

Resources

Where to find everything

Workshop landing page

neurogenomics.github.io/agentic-life-sciences-tutorial/workshops/agentic-ai-neuroscience-22-april-2026

Second Brain tutorial

tutorials/09-second-brain — full walkthrough with prompts.

Stretch goals

tutorials/10-second-brain-stretch — linting, Marp, matplotlib, search.

Go build.

Obsidian is open. Key is pasted. raw/ is empty. Drag in a PDF.
UKDRI · Skene Lab · Imperial
v1.0 · 2026-04-22