Writing a good prompt is one skill. Explaining why a prompt works to someone who has never thought in that structure before is a completely different skill, and it took me longer to get good at the second one than the first.
Part of my role involves training the trainers, meaning the people who will eventually stand in front of a room and teach prompting to ministry staff or private sector teams did not learn it the way I did, through months of trial and error against real client deadlines. They needed to learn it fast, teach it convincingly, and answer questions from a room that would test them.
The Gap Between Knowing And Explaining
The hardest part of this was noticing how much of my own prompting judgment had become intuitive rather than explicit. After enough repetitions, you stop consciously thinking through why a certain instruction works. You just know that telling a model to output structured JSON produces more reliable downstream parsing than asking for a nicely formatted paragraph, or that negative constraints tend to hold better than positive ones in longer system prompts. That knowledge lives somewhere below conscious reasoning once you have done it enough times.
The problem is you cannot teach from that place. Someone standing in front of a room needs a reason they can say out loud, something that holds up when a participant asks a follow up question. My first attempts at training material were full of instructions that were correct but unexplained, the prompting equivalent of telling someone to just trust the process.
Turning that into teachable material meant going backward through my own instincts and reconstructing the actual reasoning behind them. Why does specifying an output format reduce hallucination adjacent errors. Why does breaking a complex task into sequential stages produce more reliable results than one dense instruction block. Each of those had to become a principle with a reason attached, not just a habit that happened to work.
Building Material For People Who Will Get Questioned
Training a trainer is different from training an end user, because a trainer is going to be questioned in real time by a room, and a shaky answer in front of that room costs credibility for the whole program, not just for that one session.
That meant the materials could not just cover the how. They needed to anticipate the actual questions a skeptical participant would ask, particularly in a government or enterprise training context where people often arrive assuming AI tools are either overhyped or unreliable. Questions like why would a model ignore part of an instruction, or what happens when the knowledge base contradicts itself, or how do you know the output is actually correct rather than just confident sounding.
I ended up building what amounted to an anticipated objections layer into the training content itself, so a trainer walking into a room with a genuinely difficult audience had prepared answers rather than needing to improvise under pressure.
Solo In The Room Is A Different Kind Of Pressure
Before training the trainers became a repeatable process, some of the earliest sessions I ran myself, solo, including one focused specifically on AI for journalists. Journalists are a particularly demanding training audience, because their entire profession is built around asking pointed follow up questions and being unconvinced by surface level answers. That session forced a level of clarity in the material that the earlier internal training decks did not have, because there was no fallback if an explanation did not land the first time.
Running that session directly, rather than handing it off, ended up shaping how later training material for other trainers got structured. Every place where I had to improvise an explanation live became a place where the written material needed a clearer, more explicit version of that same explanation, so the next person delivering it would not have to improvise the same thing under the same pressure.
The Actual Lesson
Expertise that lives only in your own head does not scale, no matter how good the underlying prompting work is. The real deliverable in this part of the job was never a single training session. It was turning tacit judgment into material specific and clear enough that someone else could stand in a room, get challenged, and hold their ground using reasoning that was actually theirs to explain, not just mine.
Specific training program details, participant organizations, and internal material remain confidential given the nature of this work. Happy to discuss the general training design approach with anyone building capacity programs around AI skills through the proper channel.
Written by Mohammad Farhan Habib Faraz
Senior Prompt Engineer and Prompt Team Lead at PowerinAI
www.powerinai.com
Top comments (0)