“What is a research second brain in AI? And how do I build that?”
I got that question recently while teaching a room of 220 institutional investors.
Most investors still use AI in a fairly disposable way. They upload the documents, explain the context, and ask the question.
Then they repeat the whole exercise next week.
The hard part is keeping the research.
I have been using a different setup to scale my own investing.
And I am seeing more funds adopt it now.
A research second brain is a local knowledge base that AI keeps improving. Your original sources stay untouched while Claude turns the useful information into linked notes.
The next research does not start another blank chat. It updates what the system already knows.
So how do you actually build a second brain you can use for research?
In the next 10 minutes, I will show you my exact Obsidian setup, the copy-paste prompt, and several finance workflows you can run today.
I borrowed the basic architecture from an April 2026 post by Andrej Karpathy. His experiment had reached roughly 100 articles and 400,000 words.
Obsidian Second Brain once built out
The useful part was not the word count. It was that the research persisted.
Here is the plan:
Create a local Obsidian vault.
Give Claude access to the folder.
Paste one prompt to build the brain.
Use it on a real research assignment.
You do not need to write code or use a special database. You also do not need to lose a weekend trying to learn it from scratch.
On the Quick Start screen, choose “Create new vault.”
Obsidian calls it a vault. It is really a local folder of Markdown files, which is exactly what we want.
Click Create, name it Finance Research Brain, then choose a location you regularly back up. Choose a clear vault name and a location you can find again.
First thing you'll see is an empty vault.
Obsidian UI upon starting a new vault
Before adding AI, create one link manually.
Open a note called NVIDIA and write: "NVIDIA works with [[TSMC]] to manufacture its chips."
Click TSMC to create the second note.
That is the basic idea.
Notes store what you know, links show how the ideas relate, and sources explain why you believe them.
Step 2: Give Claude access
Download Claude Desktop, sign in, and open the Code tab.
Start a Local session and select your Obsidian vault as the project folder.
Claude Code setup
Claude’s setup sequence: Local environment, project folder, then model.
Do not use Plan mode for the build. It can inspect the folder and produce a lovely plan for files it will not create.
Send this test:
"Tell me the exact project folder and list the Markdown files you can see. Do not edit anything."
If Claude sees both notes, you are connected.
You do not need an Obsidian MCP server for this setup. Both apps already use the same local folder.
You can also do this in the terminal or with Codex. For 99% of people reading this, Claude Desktop is enough.
Step 3: Build the spine of the brain
Step 3 creates the empty system every future research job will use. Use one Obsidian vault, then give each company its own self-contained folder.
The setup prompt builds this within each company folder:
The golden architecture for your second brain!
Here is what those pieces mean:
raw/ is the evidence shelf. It holds original documents, screenshots, and source cards. Claude can read the evidence but should not rewrite it.
wiki/ is the second brain. Claude maintains linked pages about the companies, people, metrics, events, and ideas it finds across the research.
outputs/ is the work product. Initiation briefs and saved analyses live here. They can be regenerated because they are not the source of truth.
CLAUDE.md is the spine. It defines how Claude cites sources, creates pages, connects ideas, handles contradictions, and maintains the wiki over time.
You do not need to build any of this manually.
Paste this into Claude Code. When it asks what the brain should cover, type SpaceX (or whatever company you want to try first).
PROMPT.txt
SELECT & COPY
You are working at the root of my Obsidian vault. Set up a persistent company-research brain, but do not research or ingest anything yet.
Ask only: "Which company should I set up?" Resolve an ambiguous legal entity with one short follow-up, then choose a clean company folder name yourself.
Use one vault with self-contained company workspaces. Create or preserve this structure:
Never overwrite existing user content. In the root CLAUDE.md, add concise routing rules: keep company-specific evidence and synthesis inside its company folder; read that company's CLAUDE.md before working there; never mix one company's sources into another's claims; put future cross-company sector work in Industries/ and use vault-root paths for links when names could collide. Add the company and a one-line description to Companies/index.md.
Write the company CLAUDE.md as its complete research contract. Give it clear sections covering:
- Scope and ownership: the human chooses sources and questions; Claude maintains the wiki.
- File boundaries: raw/assets contains immutable originals; raw/sources contains source cards; wiki contains maintained synthesis; outputs contains dated deliverables and is not an evidence layer.
- Note contract: one durable entity, concept, metric, event, thesis, or question per note. Use YAML fields title, aliases, page_type, company, created, updated, status, and evidence_count. Start with an answer-first summary.
- Relationship rules: use [[wikilinks]] inside the company, explain each relationship in prose, add a reciprocal link when useful, and use a full path for cross-company links.
- Evidence rules: source cards record title, publisher, author and date when relevant, URL or local filename, access date, authority, exact locators, concise evidence notes, and limitations. Every material claim cites a source card or raw file plus its locator. Define one citation format, such as (Source: [[vault/path/to/source-card|Source]], p. 12), and use it consistently. Label facts, management statements, calculations, estimates, and inferences. Never silently replace conflicting evidence; show both claims, both sources, and what could resolve them.
- Navigation: index.md catalogs every wiki page with a one-line description; log.md is append-only; open-questions.md tracks gaps and next-best evidence.
- Workflows: define source ingestion, evidence-based querying, and a periodic health check for stale or uncited claims, contradictions, duplicate or orphan pages, broken links, and ambiguous cross-company references. Ingestion updates every materially affected page, with no arbitrary page quota. Queries read the index and relevant pages first, verify decision-critical points against raw evidence, distinguish fact from inference, and save reusable answers to outputs.
- Default coverage: business model, financial evidence, customers, value chain, competition, management and ownership, regulation, catalysts, risks, and thesis questions.
Finish by listing what you created and confirming that research has not started.
Claude creates the folders and writes the schema. Then it stops.
At this point, you have assembled the bookshelf. There are no books on it yet.
Step 4: Run the research workflow
Now the brain starts learning. Add a document yourself, or send Claude to research the company and follow useful rabbit holes.
Option 1: Ingest a document you already have
The original "second brain" was meant to work with documents you have already found and rely on for your research.
Start with a presentation, annual report, spreadsheet, or equity-research PDF.
Save the original file in Companies/SpaceX/raw/assets/. Then return to Claude Code and run this prompt below.
INGEST.txt
SELECT & COPY
Read the root and Companies/SpaceX/CLAUDE.md. Process @FILENAME from Companies/SpaceX/raw/assets/. Show takeaways, source caveats, and proposed wiki changes, then wait. After I confirm, follow the schema’s ingest workflow for every materially affected page. Preserve contradictions and never alter the original or another company’s workspace.
One source should update several nodes when the evidence overlaps. An earnings deck might change pages about pricing, margins, management credibility, competitors, and the thesis.
If you click on the graph view in Obsidian (second button on the left), you'll slow start seeing what resembles a small version of a neural network.
Option 2: Send it down the rabbit hole
But what happens if you want your second brain to go down the rabbit hole?
That is, for SpaceX, I want Claude to find sources, build and connect the company and concept nodes, then ingest the evidence.
The prompt reads both rulebooks and any files in Companies/SpaceX/raw/assets/, then pauses with takeaways, source problems, and a research plan.
Reply Continue and it goes to work:
Research the business model, financial evidence, customers, value chain, competitors, etc.
Update every relevant wiki page instead of producing one summary per document.
Run Prompt 1 first. This prompt reads its scope from the company-level CLAUDE.md; no configuration block is required.
PROMPT-2.txt
SELECT & COPY
Look inside Companies/ and locate the company folder created by Prompt 1. If more than one company exists and the target is ambiguous, ask me which one. Read the vault-level CLAUDE.md and that company's CLAUDE.md, then follow both. Build initial coverage only inside the selected company folder.
Phase 1: Inspect and plan
1. Inventory every file already in raw/assets/. Read each supported file fully. If a PDF is scanned, corrupted, password-protected, or unreadable, report that instead of guessing.
2. For each file, create a raw/sources/ source card that cites the original filename and useful page, slide, section, or table locators. Do not modify the original.
3. Show me one compact checkpoint containing: key takeaways from my files, source-quality concerns, likely conflicts, and the proposed public-source research plan. Ask me to reply "Continue." Do not edit wiki/ before I confirm.
Phase 2: Research and ingest
After I confirm, first integrate the approved user-supplied files across the wiki using the schema's ingest workflow. Then research up to 18 core public sources. Prioritize regulator filings, official company materials, government data, legal or regulatory records, authoritative industry sources, and credible evidence from customers, suppliers, and competitors. If the session lacks web access, stop and ask me for sources rather than fabricating research.
Cover the business model and products; customers and end markets; market structure; financial history, unit economics, and KPIs where evidence exists; value chain and dependencies; competitors and advantages; management, ownership, governance, and capital allocation; regulation and legal issues; milestones, catalysts, risks, and open thesis questions.
For every public source, create a source card with a URL and exact locator. Then integrate the evidence across all relevant wiki pages. Do not produce one isolated summary per source. Add reciprocal links, explain relationships in prose, cite material claims, preserve contradictions, and update index.md, log.md, and open-questions.md after every ingestion batch.
Phase 3: Follow the rabbit holes
After the core pass, rank the remaining gaps by how likely they are to change the investment view. Run no more than three follow-up loops, five rabbit holes, or 24 total public sources, excluding my local files. Stop when credible evidence resolves the question or more research is unlikely to change the view. Keep unresolved questions visible.
Create outputs/Initiation Brief.md and outputs/Gap Report.md. Finish with the sources used, files ingested, wiki pages created or changed, contradictions found, rabbit holes followed, remaining gaps, and the best first question to ask the brain.
LATER — INGEST ONE NEW DOCUMENT
Save the authorized file in Companies/[Company]/raw/assets/, then use this from a session opened at the vault root:
INGEST.txt
SELECT & COPY
Read the vault-level CLAUDE.md and the selected company's CLAUDE.md. Process @FILENAME from that company's raw/assets/ fully. First show me the key takeaways, source-quality caveats, and proposed wiki changes. After I confirm, create or update its source card, update every materially affected wiki page, add explained reciprocal links, preserve contradictions, update wiki/index.md and wiki/open-questions.md, and append the ingest to wiki/log.md. Never modify the original file or any other company's workspace.
The first pass uses up to 18 core public sources, followed by at most three gap loops, five rabbit holes, and 24 total public sources. If nobody tells a research agent to stop, there is always another source.
Just make sure you're checking each node along the way before letting AI loose and out of control in the wild (e.g. in your second brain).
Other workflows you can run
Public equities: Build company coverage, then update it after each earnings cycle.
Private equity: Map an industry, then extend it into target diligence.
Credit: Connect operations to cash generation, capital structure, covenants, and recovery.
Macro and multi-asset: Connect a policy shift, commodity shock, or regime to affected assets.
Technical sectors: Map product architecture, patents, standards, suppliers, and bottlenecks.
It's really up to you, and you have to set the right workflow to obtain the right data sources and research the right topics. But using this architecture levels you up from merely chatting to building a second brain.
Once the manual workflow works, create a Claude Desktop Routine that watches for new filings and earnings. That's how this grows on semi-auto mode.
Should you build this?
I think the answer is yes, if the research will recur.
Each filing and serious question improves the next piece of work. The knowledge can also compound across a team instead of disappearing when one analyst leaves.
Unfortunately, tokens aren't free.
There is always one more source to read and one more rabbit hole that sounds useful. You are paying for those tokens and reviewing the output.
Larger wikis also need better search and more maintenance. Local models may lower token costs, but they add hardware, evaluation, and upkeep.
So the rule I default to is that the value of the research brain should exceed the cost of maintaining it.
Start with one company or industry for 30 days. If the next earnings cycle or diligence sprint is not faster and better, kill it.
A second brain that does not improve your decisions is just a very organized hobby.
If you try it, hit reply and tell me what you think of it.
Enjoyed this article?
Get more AI for finance content delivered to your inbox.