AI has transformed software development, making code production an abundant and inexpensive resource. The old bottleneck of manual writing has disappeared, shifting real value toward strategic definition and critical validation. Today, the essential competency is no longer generating content, but acting as a systems manager capable of establishing solid foundations and automated evaluation criteria. Attempting to correct a "mass" of poorly structured code is far more costly than performing an accurate initial design.
Ultimately, professional success will depend on the ability to orchestrate quality processes rather than simply executing technical tasks. I was inspired by article https://dl.acm.org/doi/10.1145/3706598.3713819. Presented solution is purelly based on LLM content but could easliy extended as a coach with wearable data which will process raw metrics (like steps, heart rate, or sleep cycles) .
To illustrate this shift, two short videos are attached:
-
System A: A typical LLM chat interaction without coaching or leaderboard mechanisms.
-
System B: A well-structured environment featuring a leaderboard and When you create classes and methods "on the fly," and using metaprogramming to automate patterns to reduce repetitive boilerplate code.
The natural tendency of Large Language Models (LLMs) is to be "helpful"—meaning they default to providing answers. My Health and Wellness Coach must work against this nature: it must ask questions rather than answer them.
The brutal truth is that LLMs, despite their enormous capabilities, fall into a common trap. If we feed a prompt a "wall of text" containing 52 criteria and ask it to "rate the dialogue," the results are inconsistent. The model might forget point 14 or hallucinate point 32.
This creates an engineering gap we must bridge. On one side, we have a precise contract written in Markdown; on the other, we have raw conversation logs in JSON format. These two worlds are incompatible by default. Markdown is human-readable, but to a machine, it is just a string of unstructured characters.
Why is solving this critical for Level 4 and Level 5 systems (Contextual & Strategic AI)? Because you cannot build scalable agent systems by relying on manual log verification. Manual review works for 10 calls, but for 1,000 calls, we need automation. Without it, every change to the system prompt becomes a game of "Russian roulette"—fixing one bug might unknowingly break three other features.
- Criterion: Does the coach introduce themselves by name and explain their role in supporting health and wellness in the first interaction?
- Verification: Search for the coach's name and role declaration (coach/guide/partner) in the first statements.
- Evaluation: 1 = name + role present, 0 = at least one is missing
- Positive examples:
- "Hi! I'm Maya, your health and wellness coach."
- "Hello, my name is Leo. I’ll support you in improving your well-being."
- Negative examples:
- "How can I help?" (no introduction)
- "I am an AI assistant." (too generic)
- Criterion: Does the coach gather basic information (user name, lifestyle context, wellness goals) before deeper exploration?
- Verification: Check if the coach asked about the user's situation, habits, or goals.
- Evaluation: 1 = attempt to gather context present, 0 = coach jumps into solutions/questions without context
- Positive examples:
- "What’s your name and what area of your health would you like to focus on?"
- "Can you share what your current routine looks like?"
- Negative examples:
- User: "Hi" -> Coach: "Do you exercise daily?" (no context)
So in reality this information above in AI era is kind of openly saying YES I CAN orchestrate the process as pure programming is no longer so important.