A Claude Coach That Worked Without Me
A personalized Claude coach for each client, and a configurator that builds one. The success test was a colleague setting one up for a client with no help from me. That happened.
The problem
Training sessions end. People go back to their desks, open Claude, and forget half of it. They need help in the moment, inside their own firm's rules about what data can go where.
The coach
A Claude coach built for each client, with four modes:
- Prompt coaching. Show it a prompt, get a better one and the reason why.
- Preference setup. It interviews you and writes your personal Claude settings.
- Use-case discovery. It helps you find the work worth handing to AI.
- Onboarding. A guided first week for new users.
The method is shared across clients. The client's facts live in one file: their tools, their data rules, their approved features. Comparing four coaches line by line, most of the method files matched. Only about 14% of the knowledge files did.
It ran in live training sessions at a law firm, an apparel brand, a wealth manager and a broker-dealer.
The configurator
Building each coach by hand didn't scale, so I built a skill that builds them. It interviews you about the client, checks the facts, generates the coach, then checks it for confidentiality and features the client hasn't approved. It can also translate a Claude skill into a Custom GPT and back, or adapt an existing skill for a new client.
Ten dated versions between January and July 2026. At least eight client coaches in about two months.
The test that mattered
I set one success test: someone other than me sets up a client. In July a colleague ran the configurator for a wealth-management client. A week later he introduced the client to their coach. I wasn't involved.
What I didn't do
I never measured whether the coaches made anyone's prompts better. There was a quality checklist, but no before-and-after and no usage numbers. Next time I'd set that up in week one, before the second client.