Select your three to five best pieces of content for each format — the email you are proudest of, the LinkedIn post that got the most meaningful response, the blog post that still generates traffic — and paste them into the agent’s system prompt labelled as reference examples. Add a one-line note for each explaining what makes it work. The agent reads these examples as the benchmark and calibrates its output toward that standard.
Why Examples Outperform Descriptions
You can spend an hour trying to describe your voice in the abstract — “conversational but authoritative, direct without being blunt, uses short sentences but not choppy” — and the agent will produce something that approximates your description but misses the thing that makes your content actually sound like you. That thing is rarely describable. It lives in the rhythm, the specific word choices, the places where you break the rules.
Examples bypass the description problem. When the agent reads three emails you wrote and loved, it extracts the patterns directly — the sentence length, the opening style, the way you close, the ratio of teaching to story. It does not need you to articulate what makes those emails yours. It reads them and infers.
How to Select and Format the Examples
For each format you want the agent to produce, choose examples based on two criteria: content you are genuinely proud of (not just content that performed well — those are not always the same thing), and content that represents the voice you want to be known for going forward, not necessarily the voice of three years ago.
Format each example in the system prompt clearly. Label it: “Email Example 1 — [subject line].” Paste the full text. Follow it with a one-line annotation: “This works because it leads with a specific observation, teaches the concept through a single analogy, and closes with a question that requires a genuine answer.” That annotation tells the agent what to look for when it reads the example, not just that the example exists.
For blog posts, two to three examples are usually enough. For shorter formats like LinkedIn posts or emails, four to five give the agent enough pattern data to be reliable. Avoid outlier examples — the post that was unusually funny, unusually long, or unusually experimental. Those will skew the agent toward exceptions rather than your default voice.
What This Means for Educators
Training on past examples is the fastest path from “sounds like AI” to “sounds like me.” Most educators who have published consistently for a year or more already have excellent training material — they just have never organised it for this purpose. A one-time pass through your content archive to select six to ten strong examples, formatted and loaded into the agent’s system prompt, produces a step-change in output quality.
Update the examples every six months as your voice evolves. If you rebrand, shift your positioning, or deliberately change how you write, replace the older examples with newer ones that reflect where you are going. The agent follows the examples, so keeping them current keeps the agent current.
The Simple Rule
Show the agent your best work and tell it why it is good. Three examples per format, one annotation each. That is the fastest training investment you can make in your content production system.