A prompt scaffolding tool that helps novice users translate auditory imagination into generative AI music prompts, built from original qualitative research in a Sound and Environment course.
When students in my undergraduate Sound and Environment course were asked to generate music using AI, most of them knew exactly what they wanted to hear. What they couldn't do was say it.
resong is a prompt scaffolding tool that bridges that gap, helping novice users move from vague auditory intuition to a structured, effective prompt they can paste directly into Suno AI.
Generative AI music tools like Suno require users to describe their desired music in clear and specific language. Unlike describing an image, describing music requires vocabulary that spans genre, texture, mood, instrumentation, and acoustic quality — domains that most people have never been asked to articulate explicitly.
In my Sound and Environment course, I assigned students a creative project: use a field recording as the emotional seed for an AI-generated piece of music. The result revealed a sharp gap between what students felt and what they could write.
What patterns emerge when novice users attempt to describe music for AI generation? And what does that reveal about where they need support?
21 groups of undergraduate students submitted written prompts alongside their AI-generated tracks. Students came from a range of backgrounds. A few had formal music training, and none had prior experience with AI music generation. Their original recordings were supposed to be recordings of birdsong captured on a phone.
This context gave me a unique privilege: as the course instructor and researcher, I had deep familiarity with the assignments, the recordings, and the students' expressed intentions. I could compare what they meant with what they wrote.
I developed a five-dimension coding schema to systematically analyze the 21 submissions. Each dimension represents a distinct layer of musical description:
| Code | Dimension | Definition | Example from data |
|---|---|---|---|
| GEN | Genre | Stylistic or historical category | "jazz," "lo-fi," "mariachi," "J-pop" |
| SON | Sound quality | Timbral and textural descriptors | "high-pitched," "trilling," "staccato," "echoing" |
| EMO | Emotion / Mood | Affective or expressive qualities | "calming," "groovy," "haunting," "upbeat" |
| BIO | Bioacoustic / Sound source | References to the bird, habitat, or ecological context | "bird call," "sharp rasps," "cawing," "forest" |
| INS | Instrument | Specific instrument references | "piano," "trumpet," "acoustic guitar," "cello" |
Each submission was coded for the presence, absence, and relative weight of each dimension.
The clearest pattern in the data is that students rarely addressed all five dimensions. The typical prompt drew heavily from one or two dimensions while leaving others entirely absent.
EMO dominated. Nearly every student led with emotional or atmospheric description. Prompts like "calm, melodic smooth rhythm similar to the beat of a children's lullaby, slow tempo, and steady beat" (Group 7) and "relaxing fall music" (Group 15) convey a feeling but give the AI little sonic specificity to work with.
GEN was the most common structural anchor. Students frequently named a genre to organize their prompt: jazz, lo-fi, mariachi, J-pop, R&B trap. Sometimes with striking specificity: "J-pop, like YOASOBI's style, with some electric keyboard and guitar in the background, no jazz or heavy metal/heavy rock" (Group 5). But genre naming alone, without sonic or emotional layering, produced generic results.
BIO appeared frequently but remained underdeveloped. Some students referenced their recording by copying Cornell Lab language almost verbatim — "high-pitched metallic chips and series of loud, sweet whistles, rock piano" (Group 10) — without translating it into musical direction.
SON and INS were most commonly absent. Timbral descriptors and instrument specifications, the dimensions that most directly shape an AI's sonic output, appeared in fewer than half of submissions.
None of the three was wrong, but all three were incomplete in predictable, designable ways.
The research findings mapped directly to design decisions in resong.
The prototype below is interactive — navigate through all three screens to follow the full user flow from empty state to generated prompt.
The user arrives with their recording. The interface is intentionally quiet: a large upload zone, a prompt field with a concrete example placeholder, and the five accordion dimensions visible but collapsed. Nothing is required yet. The message is: start anywhere.
Once a recording is uploaded, the left panel contextualizes it with a waveform card and an "About this sound" box sourced from Cornell Lab. On the right, as the user writes and selects chips, the coverage bar responds in real time. The generate button remains disabled — not as a gate, but as an invitation to keep going.
The generated prompt synthesizes the user's free-text description with their chip selections into a structured, Suno-ready string. Dimension tags show which part of the prompt came from which dimension, making the scaffolding visible and educational. The user can copy the prompt, refine further, or open Suno AI directly.
resong sits at the intersection of qualitative research and interaction design in a way that is specific to my background. The five-dimension schema came directly from ethnomusicological analysis that used to study music cross-culturally and was applied to understanding how novice users describe music to machines. The design then responded to that analysis with targeted scaffolding rather than generic UX patterns.
This is what differentiates resong from a typical prompt helper: it's grounded in real data about real users, and every design decision traces back to a specific finding.
With more time and participants, I would test whether the scaffolding actually improves prompt quality, comparing AI outputs from scaffolded vs. unscaffolded prompts, rated by listeners blind to the condition. I would also explore adaptive chip suggestions based on the recording content, using audio analysis to surface vocabulary automatically. And I would design for the Feeler, Describer, and Maximizer explicitly, offering different entry paths for each user type rather than a single linear flow.