When a small investment office starts using Claude, it usually happens one person at a time. Each analyst works out their own prompts and pastes in numbers from wherever they have them. Ask two people for the same report and it comes back in two formats, and neither can say exactly which figures went into it.
I set Claude up for the whole office instead. The documents you produce over and over, such as the morning brief, investment reports, and filing summaries, become skills that anyone can run and that follow your house format. Claude gets its numbers by querying your data sources, not from pasted text. I built this for an investment manager and still maintain it.
How it works
A skill is a folder of instructions, templates, and scripts that Claude loads when someone asks for that piece of work. Writing one starts with the person who does the work today walking me through it.
Data comes through an MCP server, a small program on each machine that lets Claude query your vendor APIs and internal databases. It can read and nothing else; it has no way to place a trade or change a record. Only people who already have access to a source install the server that reaches it.
An assistant built into a research platform answers from what that platform holds. Your positions and your analysts’ notes aren’t there unless someone copies them in, and they are often what a question turns on. Through the server, Claude reads them where they already live, alongside your vendor data.
Everyone installs and updates from a server in your office, which syncs from a private repository I keep for you. A change to a skill reaches every machine from that one place.
A recorded demonstration shows the approach on public SEC filings: questions answered from the documents, with every answer tied to the passages it came from. A second, the thesis tracker, is based on my personal trading workstation. It replays an agent’s monthly runs over new export rules and four companies’ filings, and files each finding, quoted from its source, under the reason for holding the stock that it bears on.
Before anything is built
We first agree what data may go to Claude under your compliance policies. The server reads nothing outside that.
A first project covers two or three documents the whole office produces and the sources they draw on. I’ll need the person who produces each one to walk me through it, a list of your data sources and who can access each, and a machine in your office to host the update server. You supply the accounts and credentials. Strategy and model work is separate, and nothing gets write access to your systems.
When it’s done
A skill is finished when its owner has run it on real requests and the results match what they would have written. The data access is finished when each source gives the same answer through Claude as through its own interface, and the server refuses any request to change something. You get the repository, instructions for installing and updating, and a list of every source Claude can read.
After that I maintain the repository, adding and adjusting skills as your process changes, and keeping the data connections working when a vendor changes its API or Claude changes how skills run.
Message me on LinkedIn with the work your office repeats most and the data it depends on.