The 3 C's of AI prompting for financial advisors

I've been in lots of calls lately where the issue advisors are having is that they are asking a prompt to do things that they're not really saying in the prompt.

The quality of AI depends on the instructions you give it, the boundaries, and the information you're using. It's not a mind reader, and the same model can either give a really generic answer or be the best financial planner ever, depending on how well you structure that prompt.

Prompting really matters when it comes to these tools.

Frame it as delegation

There's now tools like Claude, where Anthropic's connected to the things that you're using, as well as tools like Hazel and Slant CRM. You want the prompt to be something that you can reuse and have a really usable output for you.

You need to frame this as delegation. Explain the job that you're giving the AI to do, so it has a better time making useful things for you.

There's three things that I think you want to look for when creating prompts.

Capability: assign a role

The first is capability. Assign a role to that AI model. You want it to think like a CFP professional. You want it to think like a financial analyst. Think like a professional consultant for financial advisory firms.

Constraints: say what the output should look like

First you give it the capability, then you give it the constraint.

Constraint means I want it in 150 words or less. I want this as a PDF that's one page. I don't want it to use jargon. I want this in a table, or I want this in bullet points. You give it the constraint so that you get the output you're looking for.

Context: give it the facts

Then you need to give it context: the facts. Maybe you're looking at a tax return, the goals that the client has, or client data.

You have your Claude connected to your Wealthbox, or your other tools like wealth.com. Schwab's now connected, and they're all going to be connected at some point in time.

Anthropic launched Claude for Financial Advisors in September 2026 with connectors for Schwab, Wealthbox, Wealth.com, Addepar, BlackRock, Envestnet, iCapital, Orion, SS&C Black Diamond, Vanguard, and Zocks.

That has Claude as your center point and a hub and spokes, where other tools like Hazel or Slant are the source. This is where all the meeting notes are captured, where the note-taker lives, where the emails live, where the calendar lives. It knows that already, so you're working inside that.

If you're still deciding which tools sit where in that hub, the RIA tech stack breakdown covers what belongs at the center and what belongs at the edges.

A weak prompt and a better one

Review this client data and tell me what to do next.

Review the last meeting that we had and tell me the relevant information to prompt me to prepare for this next meeting, given that they want to do Roth conversions this year.

Give it the context that it needs and give it a better version of that prompt.

Use AI to improve your prompt

You can also use AI to improve your prompt. Start by writing what you want it to do, then tell the AI, "Help me make a prompt to do this," and let it write the really long prompt where it looks at everything.

Then ask it, "Is there any information missing from this?" Keep asking it to iterate and rewrite, and it'll get better.

Save the prompts that work

When you get a prompt that you like, save it so you can reuse it over and over again. Slant makes this really easy with prompt snippets, so it lives where you're living on the client household and you can just press the same prompt: "Prepare me for this meeting," or "Review any applicable tax planning opportunities we can do this year."

Some advisors like to do observations and opportunities for prospect meetings.

Look at all this client data that's in the CRM or in this client file and prepare me for: what are the cash flow opportunities? What are the tax planning opportunities? What are the wealth transfer opportunities for this client? Give me five observations and five opportunities for each one of these areas.

It can do a really good job when you give it that context and you tell it where to look and what to do and how to make that output for you.

Do it in a system that's safe

Those three C's — capability, constraints, and context — all depend on the AI being connected to things, and you need to know that you're doing this in a safe manner. That means using enterprise agreements where there's zero data retention and zero training.

Schwab negotiated exactly that clause. Its Anthropic connector bars model training on advisor and client data at every subscription tier, not just enterprise, and masks account numbers before anything reaches Claude.

Tools like Slant and Hazel make this easy because they have the enterprise agreements, and it happens all in that ecosystem.

Don't forget that you have to disclose all this in your firm's tech policies and ADVs and wherever else that your compliance team tells you.

The SEC's 2026 examination priorities put both halves of that in writing. The Division says it will "review for accuracy registrant representations regarding their AI capabilities," and will assess "whether firms have implemented adequate policies and procedures to monitor and/or supervise their use of AI technologies." The disclosure and the written procedure are two separate things examiners look for.

Read it before a client sees it

AI's always going to make mistakes. It's not perfect, just like humans are going to make mistakes.

Always read over what it wrote before you present it to clients.

You can also ask it to give you sources of information.

Only use the IRS website for sources of information relating to gifting rules for annual gift exclusions, and use the most current 2026 number.

If you just ask it for the annual gift exclusion, it might look at 2015 and tell you fourteen thousand dollars.

That's the failure worth understanding: fourteen thousand was the correct number, for 2013 through 2017. The 2026 annual exclusion is $19,000. A stale answer that was once true is harder to catch than an invented one, which is why naming the source and the year belongs in the prompt.

Pick one recurring task

Pick one recurring task and try to make a prompt that does that. If you have a weekly sync meeting for your firm, make a task that says:

Give me a list of all the meetings I have this week, whether they're onboarding meetings or annual review meetings. Just make a list of everything so I know what's going on for the next week. I want this as bullet point output, all my annual reviews under here, all my onboardings here.

You could also have it make a table. Just try to do something where you give it the capability, constraints, and context for that task. Once a prompt earns its place in the week, it's worth writing down as a workflow with a trigger and an owner.

The best prompt's not always the longest. It just gives it the right capabilities, those constraints, and context.