The Domain Expertise Gap
The AI analyzed a track's harmony and called it 'wandering.' A musician would call it style-consistent. The data was right. The interpretation was wrong. That's not a model problem — it's a context engineering problem.
I was analyzing sections of a 6/8 meter song with a heavy waltz-like feel —
bass, piano, drums — asking a simple question:
'What's happening in this section?'
Despite tweaking the prompt over and over, feeding it the information I thought
would help, it kept reaching for vague harmonic narrative. So I stopped tweaking
the prompt and started reshaping the data — what I'd call context engineering.
Framing the prominent elements in the audio, telling the model what's in the
foreground, what to actually focus on.
The response improved. But look at where it still landed:
"Harmonically, I'm noticing some interesting little clashes... it gives the whole
section this slightly restless, sophisticated vibe... It's definitely not a
straightforward progression."
Better. But still reaching for 'restless' instead of recognizing deliberate style.
A musician hears that section and recognizes style-consistent motion. The model
referenced the chords accurately, knew about the timing, but couldn't frame the
movement as connected to the style. Accurate, confident, and still missing the point.
The model was defaulting to chord-centric analysis. Music theory text in training
data is overwhelmingly harmonic — rhythm, texture, groove are underrepresented.
Give a model raw audio features without a frame, and it narrates harmonic motion
every time. When I first measured the system's interpretive accuracy, harmonic
features dominated everything else in the signal. Accuracy was effectively zero.
This isn't unique to music. Any domain where raw data looks meaningful but requires
expertise to interpret has the same gap. The model will narrate confidently. It
won't know what matters unless you tell it.
The fix was building a focused frame for the model to look through — instead of
giving it a wall of data to make sense of. A way of saying: here's what's
intentional, what's a stylistic decision, what details in this moment actually
matter. The signal in the noise.
A musician does this without thinking — our ears sort through audio and pull out
genre, style, the function of a section. An LLM doesn't, unless you build it in.
Same model. Same audio. Same prompts. The only thing that changed was what the
model received as context. Interpretive accuracy went from zero to 75% — and
framing the foreground elements was one of the first context problems I solved.
It wasn't the last.
If you're building AI tools where domain expertise shapes the system architecture —
not just the prompt — I'd like to hear what you're working on.