4. Behavioural prompting

← Thinking in semantic roles · Guide contents · Next: Building prompts →

Most style prompts describe what a track should contain: a genre, an instrument, a production technique or a sonic texture. Behavioural prompting asks a different question:

How should the music act once those ingredients are present?

This includes how motifs develop, how parts answer one another, how energy moves through the arrangement and how performers appear to interpret the material. I use compositional prompting for the narrower part of this practice concerned with melody, harmony, rhythm and form. Behavioural prompting is the broader term because it also includes arrangement, performance and musical dialogue.

These observations come from exploratory tests in Suno v5.5. They describe what I heard, not special syntax or a verified account of Suno’s internal operation.

Exploration and specification

A broad prompt can be useful when I want surprise:

French House
Disco
Rhodes
Trumpet

Generations from this kind of prompt may be musically distant from one another. The model has plenty of room to decide what the ingredients mean and how they relate.

A more specified prompt assigns responsibilities:

Bass | Rolling Reese | Deep Sub
Harmony | Luminous Pads
Melody | Emotional Hooks
Motion | Continuous Evolution

This does not eliminate variation. In my tests, it produced different songs that remained in a more recognisable musical neighbourhood. My working analogy is the difference between rolling a d100 and rolling a d20: both involve uncertainty, but one searches a smaller space.

Neither approach is inherently better. Exploration is useful for discovery. Specification is useful when I already know which relationships I want to protect.

Production and behaviour

Some descriptors mainly suggest musical materials or production choices:

• Rolling Reese
• Ghost Notes
• Suspended Chords
• Luminous Pads

Others appear to suggest how material should develop or interact:

• Emergent Melodic Fragments
• Hidden Countermelodies
• Random Chromatic Fills
• Drifting Flow

This is not a hard taxonomy. The categories overlap, and the same phrase may be realised differently in a different prompt. The distinction is useful because it reminds me that a prompt can describe not only the palette, but the relationships and activity inside it.

Describe the audible result

Language describing an audible musical phenomenon has been more useful in these experiments than language describing an imagined implementation.

Randomly Generated Notes

was less musically legible than:

Emergent Melodic Fragments

The second phrase describes something I could listen for. It leaves Suno to decide whether the fragments belong in a voice, a bass line, a piano part or another element of the arrangement.

This is an important boundary: I can ask for a perceptible result without pretending to know the mechanism that creates it.

Context changes interpretation

Descriptors do not behave like independent switches. Their neighbours can change how they are expressed.

In one comparison, I paired:

Texture | Emergent Melodic Fragments
Motion | Continuous Evolution

The result became piano-led and melody-dominant. Its Liquid Drum and Bass identity almost disappeared, leaving only a slight two-step rhythmic trace.

I then changed only the motion phrase:

Texture | Emergent Melodic Fragments
Motion | Drifting Flow

Liquid Drum and Bass returned, the piano disappeared and the fragments often appeared as vocal ad-libs. Emergent Melodic Fragments had not selected one fixed sound. The surrounding motion language changed the musical role through which the idea was expressed.

The useful unit of revision is therefore often a relationship between descriptors, not an isolated word.

Three small observations

Random Chromatic Fills

Experiment | Random Chromatic Fills

One generation produced expressive vocal ad-libs and chromatic ornamentation. Another used chromatic bass passing notes. I did not hear arbitrary notes scattered through the track. I heard the concept placed where it made musical sense.

Hidden Countermelodies

Hidden Countermelodies did not produce two obvious, equal melodic lines. It appeared as echo responses, subtle secondary phrases and background musical dialogue. The descriptor was realised as a supporting relationship.

Emergent Melodic Fragments

Emergent Melodic Fragments carried unusually high semantic density: a short phrase substantially changed the focus and development of the composition. Its effect was powerful, but not independent of the motion language around it.

A working layered model

I currently think about these descriptors in four loose layers:

Layer	Responsibility	Examples
Genre	Establish the broad musical language	Liquid Drum and Bass, Slushwave
Production	Suggest sound, articulation and surface	Rolling Reese, Ghost Notes, Luminous Pads
Composition	Suggest musical material and relationships	Emergent Melodic Fragments, Hidden Countermelodies
Behaviour	Suggest how material changes or acts over time	Continuous Evolution, Drifting Flow, Random Chromatic Fills

This is a writing and listening model, not a hierarchy inside Suno. Its practical value is that higher-level behavioural language can help me think about how lower-level genre and production choices should be realised.

Behavioural language in a complete workflow

In Pass It On, my final Cover prompt asks vocal phrases to become synth motifs, synth rhythms to become bass articulations, and ideas to return transformed. The breakdown preserves the complete prompt and the stages leading up to it. It shows how I used that language in a track, rather than isolating the effect of any one phrase.

How I would test this more carefully

These first tests are strong enough to guide further experimentation, but not to establish repeatable rules. A more informative comparison would:

1. preserve the complete prompt, source material and controls;
2. change one phrase at a time;
3. generate several outputs for each condition;
4. record results that contradict the hypothesis as carefully as results that support it;
5. repeat the comparison when the Suno model changes.

The goal is not to discover magic behavioural commands. It is to develop language that expresses compositional intent clearly enough to support better listening and more deliberate revision.

---

← Thinking in semantic roles · Guide contents · Next: Building prompts →