Most advisors use Claude for the easy ten percent

The paste-context ceiling

Chat-and-paste has a hard limit, and you hit it around the third client.

Every session starts from nothing. You explain who the client is, what the engagement covers, what happened last month, what tone they expect. By the time the context is loaded you could have written the thing yourself, so you stop using it for anything that needs real context and keep it for generic drafting.

That is the ceiling. Not capability, setup cost per use. And it is why so many advisors conclude the tool is a better thesaurus rather than an operating layer.

What changes when it can read your files

Connect it to where your work already lives and the questions change shape. Instead of asking it to write something from a description, you ask it about your practice.

• Before. "Write a follow-up email to a client about a comp study." You get competent, generic prose you then rewrite.
• After. "Draft the follow-up for this client's comp study, in the register I used last time." It has both, because both are in the folder.
• Before. You reconstruct what happened last quarter from memory and the calendar.
• After. You ask what is outstanding across all clients, and it reads the activity logs.
• The real unlock. It stops being a place you go to write and becomes the place you go to ask what needs attention.

The control file, which is the actual trick

The single highest-leverage thing in this setup is a short standing file in each client's folder that tells Claude how to work on that client. Who they are, what the engagement covers, what is sensitive, how you like drafts handled.

It loads only when that client's folder is open, so it costs nothing until it is relevant, and it means you never re-explain a client again. One screen is enough. Longer than that and it stops being maintained, which is worse than not having one.

This is the difference between a tool that has been told about your practice once and one that knows it every time.

Structure it so the tool can find things

The thing people get wrong is bolting AI onto folders organised for a human who already knows where everything is.

The structure that works separates three things and nothing else. Standing facts about the client that rarely change. Work product, in folders named after the stages of work you actually deliver. And an append-only record of what happened, in date order.

If a file does not obviously belong to one of those three, it is almost always a standing fact. Resist inventing a fourth category. The constraint is what makes it navigable, by you and by anything reading on your behalf.

Two rules keep it from decaying: a file has exactly one home for its whole life, and a final version is marked in its filename rather than moved somewhere else. Moving files breaks every reference to them.

Where client data actually goes

This is the question that stops most advisors, and it should be asked before anything is connected.

The setup worth having runs on your own accounts and your own storage. Files stay where they already are. Nothing is uploaded to a vendor's platform, and nothing is used to train a model. You keep everything if you stop tomorrow.

Anything that requires migrating client material somewhere new is a different proposition with a different risk profile, and for regulated or people-sensitive work it is usually the wrong one.

What this does not do

It does not remove your review. Every draft is a draft. The value is starting from something specific rather than from nothing.

It does not fix disorganised files. It makes disorganisation more visible, faster. If your folders are a mess, the structure work comes first and it is not glamorous.

It does not pay off on day one. Connecting things and writing the first control file takes a session. The return starts when you stop re-explaining clients, which is usually the second week.

Questions