Prompt engineering for music AI — write prompts that get great results

The single biggest skill in AI music production is writing good prompts. Bad prompts give you generic, derivative results. Good prompts give you something you actually want to keep. Here's how to write them.

The five-element prompt

Every great prompt for music AI includes five elements:

1. Genre (lo-fi, deep house, drill, ambient...)
2. BPM (specific number)
3. Key (Am, F minor, C major...)
4. Mood / vibe (dark, uplifting, hypnotic, melancholic...)
5. Specific musical detail (instruments, swing %, characteristic moves)

Bad: "Generate a chord progression."

Good: "Generate an 8-bar lo-fi chord progression in Am at 78 BPM with jazzy 9ths and 11ths, soft humanization, in the style of Nujabes."

The good version gives the AI almost everything it needs to make a useful first take.

Patterns that work

Pattern 1: genre × instrument × characteristic move

"Generate a deep house chord progression in F minor at 122 BPM with Maj7 voicings and a mid-bar Rhodes stab on bar 3."

Specifying a *move* (the Rhodes stab on bar 3) gives the AI a concrete constraint that anchors the rest of the result.

Pattern 2: reference-driven

"Generate a drum loop in the style of Dilla's Donuts — soft kick, brushed snare, lazy swing, dusty hat."

Referencing artists or albums is one of the most effective steering tools. The AI has seen enormous amounts of training data; named references collapse a huge space of options into a useful subset.

Pattern 3: mood + technical detail

"Generate a tense, halftime trap beat at 140 BPM with sliding 808s in Cm and triplet hat rolls every 4 bars."

Pairing emotional descriptors ("tense") with technical detail ("sliding 808s in Cm") covers both feel and structure.

Pattern 4: anti-prompt

"Generate an ambient pad in C major — slow attack, no rhythmic content, no obvious melody, just texture."

Telling the AI what *not* to include is often as useful as telling it what to include. "No vocals," "no obvious melody," "no drums" all narrow the space productively.

Pattern 5: section + variation

"Take the chord progression on track 1 and write a variation that's 30% darker — same root motion, but lower extensions and minor 7th flat 5 substitutions."

For B-sections, breakdowns, and variations, anchor on the existing material and describe the *delta* you want.

Genre-specific prompt vocabulary

A vocabulary cheat sheet that's worked for us across the major genres:

Lo-fi

• "swung," "dusty," "warm," "tape-saturated," "vinyl crackle," "lazy"
• Common keys: Am, Cm, Em, Dm
• BPM: 70-90

Trap / drill

• "808 sub," "sliding pitch," "triplet hats," "halftime kick," "dark"
• Common keys: Cm, Dm, Fm, F#m, Gm
• BPM: 130-160

House / deep house / tech house

• "four-on-the-floor," "off-beat hat," "shuffled," "sidechained," "Maj7"
• Common keys: Am, Cm, Dm, Em, Gm
• BPM: 118-128

Techno

• "driving," "hypnotic," "modal pad," "303 acid," "pulsing bass"
• Common keys: Am, Cm, Dm, Fm, Gm
• BPM: 125-140

Ambient / cinematic

• "evolving," "drone," "atmospheric," "modal," "long reverb tails"
• Any key works
• BPM: 60-120

What to specify per task

For chord progressions

• Number of bars (4, 8, 16).
• Key.
• Chord types (triads, 7ths, 9ths, sus chords).
• Voice leading character ("smooth," "jumpy," "modal").
• Tension/release pattern.

For drum loops

• Genre + BPM.
• Time feel (straight, swung, halftime, double-time).
• Specific moves (fill on bar 7, ghost notes on the snare).
• Sound character (acoustic, 808, drum machine, trap, house).

For basslines

• Following the chord roots vs walking vs syncopated.
• Sub vs mid bass.
• Sidechain to the kick (yes/no, intensity).
• Movement (static, octave-jumping, syncopated 16ths).

For melodies

• Length in bars.
• Range (octave, two-octaves).
• Rhythmic character (busy, sparse, syncopated, on-the-beat).
• Phrasing (call-and-response, statement, hook).

Common mistakes

Too vague

"Make me a beat."

Useless. The AI's only choice is to give you the genre-average. You'll get something boring.

Too constrained

"Generate a chord progression that goes Am, F, Cmaj7, G with a half-bar pickup, ending with a deceptive cadence to D minor 9 sharp 11..."

You've already written the song. Just write it. AI helps when you give *direction* without writing the answer.

Mixed-genre confusion

"Generate a lo-fi chord progression with hard trap drums and ambient pad textures at 175 BPM in F# major with vocal chops."

The AI will give you a confused mash-up. Pick one genre per prompt; combine results across multiple prompts in your DAW.

No reference for novel asks

"Make me something that sounds completely new."

The AI is interpolating its training data. It can give you variations on existing music; it cannot give you music that has no precedent. If you want truly new sounds, use AI for the conventional parts and bring your own ideas for the novel ones.

Iteration is the real skill

The producers getting the best results from AI aren't the ones with the perfect first prompt. They're the ones who iterate fast.

A typical iteration loop:

Prompt 1: "Generate a lo-fi chord progression in Am at 78 BPM with jazz 9ths."
→ Result: OK but predictable.

Prompt 2: "Same key and tempo, but with a deceptive cadence and a borrowed chord from the parallel major."
→ Result: more interesting harmony.

Prompt 3: "Same progression but voice it lower and add Maj7#11 instead of Maj7."
→ Result: that's the one. Save it and move on.

Three iterations, less than two minutes, and you have something better than your first prompt would have produced.

Tools where prompt engineering matters most

• VIXSOUND (chat-driven, prompt is the entire interface).
• Suno / Udio (prompt is the only knob you have).
• MusicGen / MusicLM (research models that respond strongly to prompt wording).

Tools where prompt engineering matters less:

• Captain Plugins / Scaler (rule-based, GUI-driven).
• Output Arcade / Splice (browse-based).

Read next

• How to use AI in Ableton Live
• AI MIDI generation explained
• Ableton AI workflow from scratch

Summary: be specific, use reference points, iterate fast. The producers who treat prompting as a craft (the way developers treat prompting Cursor or Claude) get noticeably better results than the producers who type whatever comes to mind.