Prompt repair: what to do when an AI music generator ignores you
When a generator ignores an instruction, adding more instructions is often the worst response. Long prompts can contain conflicts, repeated ideas and low-priority details that compete with the part you actually care about.
Reduce the prompt to three non-negotiables, one arrangement move and one avoid instruction. Test that version first. Add detail back only after the core behavior appears.
Find contradictions before blaming the model
Read the prompt as if you were a producer receiving a brief. “Minimal but huge,” “slow and frantic,” or “dry vocal with massive reverb” may be artistically possible, but they need section context. Without it, the generator must guess which instruction wins.
- Separate section-specific instructions from whole-song traits.
- Remove synonyms that repeat the same idea.
- Do not specify five genres unless you explain what each contributes.
- Check whether the avoid list contradicts the desired style.
Use the compression test
Rewrite the prompt in one sentence under roughly 35–50 words. If the song direction disappears when you shorten it, the original prompt probably depended on decoration rather than decisions. Keep the compressed version as the control test.
Move priorities to the front
Put the most important musical behavior first: groove, vocal character, core palette or arrangement change. Secondary mix adjectives and optional details can come later.
Emotional cinematic atmospheric modern dark polished song with many details... and somewhere in verse two use double-time vocals.
Mid-tempo dark alt-pop; verse 2 switches briefly to double-time vocal phrasing over the same BPM; low intimate lead; sparse drums and warm sub; chorus opens wider; no EDM drop.
Change one variable between tests
If you rewrite genre, voice, structure and production together, you cannot know what fixed the result. Keep a small prompt log: version, one change, what improved, what regressed. After three or four tests you will have evidence instead of guesses.
Repair order: remove contradictions → compress → place priorities first → test → change one variable → add detail back only when needed.