The single biggest predictor of whether a generated song sounds good is the prompt. Lyria 3 will follow a vague prompt vaguely. It will follow a specific one specifically. After generating a few hundred tracks on Brainy Records, here are the patterns that consistently produce songs people actually finish listening to.
The general shape
Every prompt that works has the same skeleton: genre + reference acts + production cue + arrangement note + mood. In that order, in one sentence each, ideally under 250 words total. Less prompt = more model latitude. More prompt = closer to your taste.
A bad prompt looks like: "a pop song about love." Lyria will produce something — usually generic mid-tempo pop with a vague chorus.
A good prompt looks like: "Stadium pop with anthemic choruses, in the style of late-2010s OneRepublic crossed with The 1975. Polished studio mix, layered vocal harmonies, a four-on-the-floor kick under a synth lead. Three distinct sections — intro, big hook, breakdown back into a final hook. Mood: confident, slightly bittersweet, drive-music energy." Now Lyria has constraints to work inside, and the output will reflect them.
Metal
Metal is one of the model's strengths — surprisingly. The melodic core of metalcore and modern metal survives quantization better than, say, acoustic folk. What works:
Reference acts that produce coherent output: Sleep Token, Bad Omens, Bring Me The Horizon, Lorna Shore, Polyphia, Periphery. These all have studio-polished production that Lyria can imitate. Reference acts that struggle: anything heavily improvisational like Mastodon's longer tracks, or anything with deeply weird vocal tone like Cattle Decapitation.
Production cues that work: "down-tuned guitars, polished djent production, layered clean vocals over harsh shouts in choruses, programmed kicks." Avoid asking for "raw" or "lo-fi" — Lyria will deliver a mush. The model is better at clean than at lo-fi for heavy music.
Arrangement: ask for a melodic verse, a heavy breakdown, and a clean reprise. That three-part structure consistently lands. See what's in the metal genre on Brainy Records for examples.
K-pop
K-pop is the other strong category, for opposite reasons: it's already a polished, highly produced sound, which is what Lyria does best.
Reference acts: NewJeans, ITZY, ATEEZ, IVE, Stray Kids, LE SSERAFIM. These have distinct sonic identities the model can lock onto. Avoid asking for "an authentic Korean group" — that's a meaningless instruction; ask for the sonic and structural qualities you want instead.
Production cues that work: "vocal stacking, half-time hip-hop drum drop in the bridge, bright synth hooks, side-chained pads, English-Korean code-switching in the lyrics." Lyria's vocal generation is decent enough for k-pop hooks but struggles with rapped verses — accept this and structure prompts around sung melodies.
Arrangement: k-pop loves the post-chorus drop. Ask for "a chorus that resolves into a wordless post-chorus hook, then a half-time bridge before the final chorus." See the k-pop genre.
Hip-hop
The model's weakest mainstream genre. Lyria 3 generates passable rap vocals on simple flows, but anything intricate (Kendrick, MF DOOM-style internal rhyming, complex pocket work) doesn't survive. Trap and lo-fi hip-hop are the safer bets.
Reference acts that work: Tyler the Creator (Igor era), Mac Miller (Circles era), Brockhampton, Daniel Caesar's hip-hop adjacent stuff. These all have melodic vocal performances that Lyria can imitate.
Production cues that work: "808s with a slow attack, dusty drum machine kit, jazz piano sample chopped up, ad-libs scattered through the second half." Asking for "boom bap" produces something — a stiff loop. Ask instead for the textures: "warm vinyl crackle, swung kick-snare pattern, mellow horn samples."
What doesn't work: aggressive trap with rapid-fire 16th-note hi-hats. The model will produce something but the hats sound wrong. See /genre/hiphop.
Synthwave
Best results come from explicitly visual prompts. Synthwave lives or dies on aesthetic, and aesthetic translates into specific instrument choices.
What works: "1984-coded synthwave with analog Juno-60 pads, a Linn-style drum machine, arpeggiated bass synth, sparse vocoded vocal layer." Reference acts: The Midnight, FM-84, Carpenter Brut, Mitch Murder, Perturbator.
The model handles instrumental synthwave better than vocal synthwave. If you want vocals, ask for "thin reverb-soaked vocal melody, low in the mix, evoking a memory rather than fronting the song." See the synthwave bands.
Hyperpop
Counterintuitively easy. Hyperpop's core aesthetic is "everything maximalized at once," which is what happens by default when you over-prompt the model. Don't worry about restraint here.
Reference acts: 100 gecs, A.G. Cook, SOPHIE, Charli XCX (Brat era), Underscores. Production cues: "brick-walled mix, pitched-up vocal layer, glitching synth bass, breakcore drums in the bridge, intentional digital artifacts." Lean into the genre's noise and the model leans with you. /genre/hyperpop.
Reggaeton
Lyria does dembow well. Reference acts: Bad Bunny, Rauw Alejandro, Karol G, J Balvin (mid-period).
What works: "perreo-tempo dembow riddim, side-chained reggaeton kick on the half-bar, melodic Spanish vocal hook, brass stab as transition between chorus and verse." Avoid asking for "Latin music" — too vague. Ask for dembow, perreo, or specifically reggaeton. /genre/reggaeton.
Things that work in any genre
Specificity beats verbosity. A 200-word prompt with three vague style descriptors loses to a 50-word prompt with three specific reference acts and one specific production cue.
Anchor the song's structure. Even one sentence about arrangement ("intro builds for 8 bars then drops into the hook, second verse is half-time, bridge is acoustic, final chorus drops the bass and brings it back") keeps the output coherent for the whole length.
Ask for one signature element. "A wailing slide guitar that returns in every chorus" gives Lyria something to anchor the track around. Without an anchor, the model produces music with no memorable through-line.
Match production cues to era. "Modern radio polish" produces a different song than "warm 1970s vinyl mix" or "cassette-tape lo-fi room mic." Pick one and commit.
Things that consistently don't work
Asking for the song to "tell a story" — the model doesn't do narrative coherence in lyrics yet. Ask for a feeling instead.
Asking for "10 minutes long" — Lyria 3 Pro Preview targets ~90 seconds. Longer renders just truncate.
Asking for "live recording feel" — the model produces studio mixes by default and converting that to live-room ambiance is unreliable.
Asking for a specific instrument by brand AND time period AND playing style — pick one or two of those. Three is over-constraint and the model gives up and goes generic.
Try a prompt right now
If a few of these patterns gave you a song idea, generate it on Brainy Records. The first vocal song is free per Google account, and the prompt becomes part of the band's profile so others can see what worked.
Or browse /discover — every band there has its style description visible, which is roughly the prompt that produced it. Reverse-engineer the ones whose songs you like.