Why not just ask Claude?

TL;DR: A strong model with your vault open will write you a good strategy memo in ninety seconds. Then it's a note, and the next thing you ask can't tell whether that note is still the company's position or one you moved past. StratOS keeps the decision in a shape that can be wrong, and your agents read it before they build.

You should

The question arrives in the first five minutes of most conversations, and it's sharper than "why not a chat window". The founders asking it already run Claude Code with an Obsidian vault open, a folder of skills they wrote or pulled from skills.sh, MCP into Notion or Linear, and, for the earliest of them, a fast structured model such as TypeSafe's Jev, in early access since September, for the typed calls inside the workflows. That stack is what we have to compare against, because it's what the founders we talk to run.

You should use it. Ask Claude Code, with the vault open, "where do we stand and what should we bet on next quarter", and in ninety seconds you get a memo a consultant would have taken two weeks over: stage, customer, three recommendations. We build on the same models for exactly that reason.

The answer isn't that our model is better. It's what happens to the memo after it's written.

What the vault holds

The memo goes into the vault as a dated note. Next month's positioning note links to it. The month after, you change the segment in the customer note. The memo and the positioning note still say the old thing. Backlinks tell you the notes are connected; nothing tells you they're now wrong.

Claude Code, reading the vault for the next task, finds the old memo and the new customer note with equal authority, because to an agent a note is a note. The pricing skill you wrote reads whichever notes you remembered to point it at, and it can't tell a claim you verified from one the memo assumed, because a note has no grade.

Concede it honestly: a vault that keeps a decisions folder, each note dated, with what it rests on, the number that would retire it and the date to re-read it, re-read on schedule, with the dependents marked stale manually when one changes, is the record built by hand. Some founders keep one. Most vaults we see hold the memo, not the decision.

The gap isn't what the vault can store. It's what nothing in the stack enforces: the shape of a company decision, the grade of a claim, and what goes stale when something upstream moves.

The first read

Our most important read is the first one: where the company stands. It starts from your URL and what you connect.

Before the model writes a word, the server builds the frame from your evidence. The frame names the one problem that most changes how your current growth should be read, the competing interpretations of that problem, and the evidence that would settle between them. The model elaborates that frame. It can't swap in a more interesting one.

It's also bound by four rules a memo isn't. One of them: it may not state as fact the thing its own bet is designed to test. And it opens on six fixed answers instead of fifteen hundred words. All four rules and the six answers are written up as a test you can run on your own memo, in Open your repo and try to find a decision.

Then it's committed. Every later read, whether positioning, pricing, or the spec, starts from that, not from whichever notes the agent happened to open.

What the record keeps

The server owns the shape of every output. A skill can't name a competitor that was never served, state a figure with no source, cite a framework to sound authoritative, or restate a committed number with a different value. Gates on every run, not instructions.

Every output is a position with a threshold, a kill criterion and a read date. Nothing enters without your commit. A committed decision is never rewritten; it's superseded with the cause. When the first read changes, the reads that rested on it are marked stale with the reason, by the record, not by you remembering.

When a number crosses a threshold, the record asks you for a verdict and never gives one itself. What is waiting for a verdict sits on the board until you give one.

And then your agents read it

StratOS doesn't compete with Claude Code. It's what Claude Code reads. Keep the vault for what it's good at.

An agent connects with one key and gets the committed strategy: the current positions, the open bets with thresholds, the moves you sanctioned, the build spec for a live decision. For a superseded decision it gets a refusal before any prose, so it doesn't build the version you moved past. It can write back evidence and can't approve, resolve or change terms.

What the refusal buys is timing: you learn the decision is dead before the agent writes the first file, not in review.

So ask Claude. With the vault open. Then give it the record to read first, so it knows which of the notes in your vault the company is actually betting on.

A model is the expert in the room for ninety seconds. The vault is what you wrote down. The record is what the company decided.