Most firms cannot describe their ideal client past "small business owner who pays on time." This is the whole exercise, start to finish. It takes about an hour, it needs no client data, and at the end you have something that feeds your proposals, your pricing, your marketing, and what you say no to.
This is a worked example from our AI for Tax and Accounting Pros guide. It is the one we recommend people run first, because it involves no client data and no tax law, and it is the kind of work most firm owners never make time for.
Why this one first
Everyone reaches for tax prep and workpaper automation first. Those are real uses, and they are also the ones carrying the most client data and the heaviest review burden. This exercise runs on characteristics rather than identities, so you get the full value of the tool with none of the data question in the way.
It also uses all four levers at once: real context only you have, standing instructions that persist, memory carried across sessions, and back-and-forth to sharpen the result.
The exercise
1. Open a Project
Call it something like "Ideal client." Everything below lives inside it, so each step builds on the last instead of starting over. In Claude this is a Project. Other tools have an equivalent container.
2. Describe your ten best clients
For each one: industry, size, entity type, what you do for them, what they pay, how they found you, and why they are easy to work with. Then describe three you would hand back tomorrow, and why.
Write these as types, not people. "Manufacturer, four million in revenue, S-corp, three entities, owner is decisive and answers email" describes a type. It is the pattern you are after, not the name.
3. Ask it to find the pattern
This is the part you cannot do alone, because you are too close to your own book.
Here are ten clients I love and three I would hand back. What do the good
ones share that the difficult ones do not? Group the patterns. Then tell me
where I am probably wrong about my own book, and what I seem to be
optimizing for without saying so.Push on the answer. Ask which pattern is strongest, which is coincidence, and what it would take to disprove each one.
4. Turn it into a written profile
Ask for a one-page profile with explicit disqualifiers, not just an ideal. The disqualifiers are the half most firms skip, and they are the half that saves you time.
Then save that profile into the Project instructions, so every future conversation in the Project already knows who you are trying to serve.
5. Now screen against it
Here is what a prospect's website says. Score the fit against my profile,
name the mismatches explicitly, and draft the three questions I should ask
on the call to test the biggest unknowns.That is the payoff. The profile stops being a document and starts being a filter you run prospects through.
What you end up with
A written profile with disqualifiers. A repeatable screen for new prospects. Three call questions that are actually about fit rather than about scope. And a standing instruction in the Project, so the next thing you draft in there already knows your ideal client without being told.
It compounds from there. The profile feeds your proposals, your pricing conversations, your marketing copy, and the work you decline.
A note on data
This runs on characteristics, not identities. You are describing types of businesses and types of owners, not naming taxpayers, so the whole exercise stays clear of client information. If you find yourself pasting in a name, a return, or a set of financials, stop and generalize it first. The exercise works better that way anyway, because patterns are what you are looking for.