Music Generation Gets the Same Skill Pattern as Code

lyria, at 9.3, is one of the highest-scoring single skills in the directory — and it generates music. The interesting part isn't that an agent can make audio. It's that the skill earned its score the same way the best coding skills do: by closing the loop between what was asked for and what actually came out.

When people say "agents can only do text-shaped work," music is the supposed counterexample in both directions — either obviously impossible, or, now that audio models exist, trivially solved by asking. lyria — 9.3, wrapping Google's Lyria 3 music generation — is more interesting than either claim. It treats music generation as a software-style problem: specify precisely, generate, then verify the artifact against the spec.

What the skill actually does

Three layers, in the order that matters:

Prompt construction. Natural-language prompts to a music model are notoriously lossy — "a melancholy piano piece with some strings" gives the model almost nothing to hold onto. The skill encodes musical intent in structured terms: genre conventions, tempo and key constraints, section-by-section structure (intro, verse, bridge), instrumentation, mood targets. This is the same move a good coding skill makes when it turns "build a REST API" into a concrete contract — name the dimensions, then constrain them.

Structure control. Song form is the part naive prompting loses entirely. Without explicit section-level scaffolding, generative audio drifts — the "chorus" isn't a chorus, the energy arc goes flat in the middle third. The skill's structure pass is the musical equivalent of architectural review before implementation.

The verification loop. This is where the 9.3 comes from, and it's worth dwelling on. Generated audio sounds like music almost unconditionally — the failure mode is music that doesn't hold together: out-of-key moments, dropped tempo, a section that references nothing and goes nowhere. The skill defines checks for exactly these: does the piece stay in the intended key, does the tempo hold, do sections progress like sections, does the track resolve. The generation step can't confirm any of this on its own; the verification step can, and lyria runs it before returning the artifact as done.

Why verification is the whole story

Compare with the audio-skills-that-didn't-score pattern: prompt templates that fire a generation call and hand back whatever arrives. Those skills have the same model underneath and dramatically worse outcomes, because an unverified music artifact fails silently — the requester hears ten seconds of plausible sound and only discovers the structural collapse after they've already built on it.

Silent failure is the common enemy of every generative domain, and the skill ecosystem has been converging on one answer regardless of medium. Code skills add "run the tests." Image skills add "check against the reference" — that's precisely the extract-then-reapply contract behind visual-style. Asset-hygiene skills like baoyu-compress-image go a step further and treat format and quality as measurable thresholds, not taste. lyria adds "confirm the music doesn't fall apart." The pattern is medium-independent: generation gets you an artifact; verification gets you a deliverable. The 9.3 is for knowing the difference.

The domain argument

There's also a quieter point about domain boundaries. "Agents do text, humans do music" was always a statement about interfaces, not capabilities — the music domain just lacked the contract structure that code has had all along. The moment a domain acquires machine-checkable properties (key, tempo, section logic, duration), it acquires the same skill anatomy: specify, generate, verify. Music turned out to have those properties; color and duration and structure are measurable facts, not opinions.

So the score isn't a curiosity — it's a forecast. Any creative domain with verifiable structure will follow this exact template: tight prompt contracts, structural scaffolding, a verification loop that keeps the agent honest. Music is simply one of the first to get it.

An agent that can generate music isn't the news. An agent that can tell whether the music it generated is any good — that's the skill, and that's the 9.3.