Use source material that contains complete reasoning: client question, missing context, your diagnosis, the options you considered, and the next step you chose. Add examples of when your usual advice does not apply. Then test for false confidence, agreement bias, and invented claims before launch.
Catch yourself making a real decision
The most useful training clip is often not the keynote. It is the five minutes after a client says, “I tried that and it did not work.” You ask two questions, discover the real constraint, and change the recommendation. That sequence contains your method in motion.
Save examples with friction: the client who wants permission, the plan that looks right on paper, the exception to your own framework. Smooth teaching explains the rule. Difficult coaching reveals how you use it.
Write down the questions you ask before answering
Generic AI tends to answer too early. Coaches often earn their fee by refusing the first version of the question. “Should I lower my price?” may actually mean “I do not know how to explain the offer,” “I need cash this week,” or “one prospect said no.”
Create a short diagnostic list for each major topic. What must the coach know before advising? What evidence would change the recommendation? Which answer sounds reassuring but creates a worse problem next month? Those questions give the AI a better route through the conversation.
Train the edge of the map
A trustworthy coach needs to know what it cannot coach. Add explicit examples for mental health crises, medical concerns, legal disputes, investment decisions, threats, abuse, and any regulated advice outside your credentials. The response should acknowledge the person, name the limit plainly, and point toward appropriate human help.
Also define ordinary commercial boundaries. If your group program is not for beginners, say so. If you never advise someone to quit a job from one conversation, teach that restraint. Boundaries are part of the voice because they show what you take seriously.
Correct the answer in your own words
When a test answer misses, do not write “be less generic.” Replace it with the answer you would actually send. Add the question you would ask first. Remove the promise you would never make. Point to the lesson the system should have used.
A series of concrete corrections becomes a second, sharper training set. It captures the difference between language that merely resembles you and judgment you would put your name behind.
Keep a simple correction log: the question, the bad answer, why it failed, and the version you approved. After twenty entries, you will see recurring gaps in the source material and recurring habits in the model. Fix the pattern, then rerun the old questions.
Questions for every test answer
- Did it ask for context before giving consequential advice?
- Can I point to the source behind the recommendation?
- Would I send these words to a paying client?
- Did it invent a story, statistic, credential, or promise?
- Should this conversation move to a human?
What coaches usually ask next
Can an AI copy my exact voice?
It can learn patterns from your material and produce responses shaped by your language and method. You still need to test and correct it; no upload guarantees perfect imitation or judgment.
Should I upload private client calls?
Only if you have the rights and consent required to use them. Safer alternatives include your public teaching, anonymized case material, role-played examples, and new recordings made specifically for training.
What if my older content contradicts my current teaching?
Leave it out or label the change explicitly. A smaller current corpus is better than an archive that presents both versions as equally valid.
How often should I retrain it?
Add or revise material when your method, offer, scope, or policies change, and whenever repeated client questions expose a real gap.