Research that compounds
into an edge.
Laniakea Brain is a research system built on Claude Code and Obsidian, installed on your own machines and written around your process. We run it on our own book.
What a new install looks like.
One folder on your machines, one editor, one engine, and a sensing layer that runs overnight. Everything below is plain text you can open and read.
| Piece | What it is | What we set up |
|---|---|---|
| The vault | A folder of markdown files, opened in Obsidian: theses, research notes, sector maps, macro notes, mental models and a few state files | Folder structure, templates and the rulebook, in your house style |
| The engine | Claude Code, reading and writing the vault from a chat panel inside Obsidian | All 27 skills and 8 workflows, adapted to how your desk already works |
| The sensing layer | Scheduled workflows that watch news, prices, catalysts and X between sessions and write into a scanning folder | Tuned to your names and themes, with alerts to your phone |
| The team | You choose the sources, argue in the margins and take every conviction decision | Trained to direct, verify and argue with the draft |
Nothing is hosted by us. If we disappeared tomorrow, the vault would still open.
Raw sources, the wiki, the schema.
| Layer | Contents | Who writes it |
|---|---|---|
| Raw sources | Inbox deposits, overnight news and X sweeps, market data, your own objections in the margins | The world and you. The model never modifies a source. |
| The wiki | Your theses, sector maps, macro notes and research notes, plus the state files: session memory, dependency map, follow-up register, catalyst calendar | The model, through named skills |
| The schema | A rulebook, the note templates, the skill specifications | You, rarely |
Sources never change, so provenance survives every rewrite above them. The schema keeps every machine-written file the same shape, which is what makes them readable as a set.
Chat is not a research system.
| Chat window | Search over your files | Compiled research wiki | |
|---|---|---|---|
| Memory between sessions | None | An index. Nothing is learnt. | Written back into the notes |
| Cross-references | None | None | Built at write time, kept current |
| Contradictions | Unflagged | Unflagged | Checked every time a source is filed |
| Audit trail | None | None | Dated, append-only log per position |
Compilation, not retrieval.
Six engines.
Three interrogate your book, one audits conviction, one ingests the outside world, one compiles and publishes.
| Engine | What it does for your book |
|---|---|
| Retrospective | Reads what you wrote against what the market did, and ranks the names where the two came apart. |
| Cross-thesis | Scans every thesis at once for shared dependencies, contradictions and stale assumptions. |
| Adversarial | Attacks a thesis as a hostile short-seller and writes down the break paths. |
| Conviction audit | Checks each position against the raise, cut and close triggers you registered in advance. |
| Ingestion | Turns a source into a structured note and propagates it to every file it touches. |
| Publishing | Compiles a memo or brief from the notes, with every claim traceable. |
The engines write drafts. They never write a conviction.
Five ingredients in every prompt. Three of them compound.
| Ingredient | What it is | Who writes it |
|---|---|---|
| Typed command | What you type in the box. Different every time. The only part most people call "the prompt". | You, each time |
| Conversation history | Everything said earlier in the session. Comes free with the model. | Free, uncontrolled |
| Context files | The model's memory and rulebook. Written once, read at the start of every session. | You, written once |
| Templates | The shape the answer must take. The model cannot skip the bear case, because the bear case is a section. | You, written once |
| Skills | Named procedures the model follows step by step: ingest, propagate, stress-test. | You, written once |
Prompt quality is mostly what the model already knows before you type. Ingredients three, four and five are where the work goes, and they compound.
What it knows before you type.
| File | What it holds | Why it matters |
|---|---|---|
The rulebook CLAUDE.md | Vault structure, writing standards, decision rules. Loaded every session. | Every answer inherits the house rules. |
Session memory _hot.md | The active thread, the latest sync, recent conviction changes, open questions. | The model resumes mid-thread, not from zero. |
The dependency map _graph.md | Which theses link to which sectors, peers and evidence. | It opens the handful of files that matter instead of re-reading the book. |
The follow-up register _followups.md | Findings that demand action: stress-test flags, fired triggers. | Nothing auto-forgets. Entries leave only when resolved. |
Mental models Mental Models/ | Your firm's own lenses, written down as questions, never conclusions. | The house edge applied on every pass, not when someone remembers. |
When every lens likes the same name, that is the cue to argue the other side.
Templates and skills. Judgement into procedure.
Templates fix the shape
Every thesis follows one template: summary, non-consensus insights with their falsifiers, bull case, bear case, risks, conviction triggers (raise if, cut if, close if), and an append-only dated log. Every research note follows one template too, leading with what changed for the thesis and a mandatory contradiction check. Identical skeletons can be compared against each other. Bespoke documents cannot.
Skills fix the method
A skill is a written procedure the model follows step by step: pre-flight checks, method, write discipline, then a report of every file it touched. Same command, same shape of output, whatever the model's mood. Eight portfolio-wide workflows fan one skill across every thesis at once, and sceptic agents attack the findings before you see them. None may touch a conviction.
The 27 skills. All of them ship with every install.
● The core loop
◆ Analytical
▲ Building
◐ Diagnostic
■ Publishing
# All output is canned and deterministic, same skill, same trace.
# No API calls. No live data. Fictional ACME, Acme Robotics.
The same question, typed into both.
“What did this morning’s news actually touch?”
Bare model
A tidy summary of the headlines. No idea which names you hold or why. Cannot say which claim in which thesis changed. Gone by tomorrow; the context has to be rebuilt.
Vault-loaded model
Reads session memory, the scored brief and the dependency map first. Names the theses touched, and why each. Tests each story against the thesis's own conviction triggers. Appends a dated log entry per thesis, linked to the source.
Same model. The difference is what it already knew, and what it was allowed to write.
The sensing layer.
A vault only knows what someone put in it. Between sessions nothing watches the observables your theses depend on. Five scheduled workflows fix that. They watch. They never judge.
| Does | Never does | |
|---|---|---|
| The sensing layer | Pulls on a schedule, removes duplicates, scores for relevance, sends alerts | Judgement. It has no mental models and no thesis state. |
| The vault skills | Analysis, propagation, conviction | Watching the world between sessions |
| You | Pick the one to three stories worth the book's attention | Read thousands of items |
- OvernightNews sweep. Several channels read in parallel: hand-picked outlets, news search, a financial-data feed. Thousands of items deduplicated, scored, clustered, the best re-read in full and summarised. You see a brief by theme and ticker, and the top stories on your phone.
- Early morningCatalyst reminders. Alerts two days out and on the day for every dated event in your book.
- Early morningPrice tripwires. Each thesis's raise, cut and close levels checked against the close. An alert quotes the trigger it threatens.
- Before the desk sits downX harvest. Posts on your names and themes tracked over days and read against your own bull and bear cases. Where the crowd argues something your thesis does not carry, you hear about it.
Nothing from this layer enters the book by itself. It writes only into a scanning folder, and it costs $20–35 a month to run.
Questions you would ask if a friend pitched this to you.
Do I need to be a programmer?
No. We install and configure everything: Obsidian, Claude Code, the vault, the templates and every skill, written around your process. Day one, you are calling skills, not setting them up.
You do have to be comfortable typing a command instead of clicking a menu. If you live in spreadsheets, you will be fine.
Will the model write things into my thesis that are not true?
Treat every section it writes as a first draft. The source is locked in the note's frontmatter, so any claim traces back. Inline callouts let you push back on a sentence and keep the disagreement on the record.
What stops it going off-script and breaking the vault?
The rulebook sets the conventions, and each skill is a step-by-step specification rather than a free-form prompt. Conviction shifts, status changes and deletions need your confirmation. Destructive operations snapshot first, and a health check audits drift.
How long until it is useful?
The first thesis is already worth more than ten loose notes. The compounding shows up around the fifth or sixth, the first time a sweep finds a pattern across positions you could not hold in your head. Most desks feel it after a month of regular use.
Where does my data live?
On your machines, as plain markdown you own and version yourself. Nothing is hosted by us, and nothing is used to train anyone's model.
Is this only for equities?
The templates and skills are written for investment research, but the pattern generalises to any long-running research work. Same primitives, different templates.
The eight features that carry the return.
Compilation, not retrieval
Knowledge written back into the notes. March’s insight cited in September.
Interlinked context
Every note knows its neighbours. A datapoint finds the handful of files it touches.
Pre-registered triggers
Raise, cut and close conditions written before the market has an opinion.
Propagation
One source updates every affected note in one pass, with dated log entries.
Adversarial procedures
Stress tests, scenarios and sweeps whose purpose is to attack the book they maintain.
A register that never forgets
Findings that demand action persist until resolved.
Callouts
Disagreement preserved next to the claim it concerned.
The sensing layer
Thousands of items a night reduced to a brief. Divergence flags against the crowd.
None of them is the model. All of them are what surrounds it.
Have us install it on your stack.
We run Laniakea Brain on our own book, and we build it for other funds.