Claude Opus 5.5 Composed Retro Game Music From a Single Prompt — Here Is What That Means
Simon Willison's Scrimshaw Jukebox experiment shows Claude Opus 5.5 designing its own music notation format, building a browser-based player for it, and composing retro-style game music — all from one prompt. The capability is real but unverified at scale, and key limitations around audio export, format portability, and model-tier access mean non-technical users should approach it as a promising prototype, not a finished workflow.
A Single Prompt Produced Playable Game Music
Something worth paying attention to happened quietly in early October 2026. Simon Willison, a well-known developer and AI researcher whose work is closely followed in the developer community, published a short but significant experiment on his site. The post, titled Scrimshaw Jukebox, documents what happened when he gave Claude Opus 5.5 a single instruction: design a simple text-based format for music, build an artifact that can play it out loud, and include some example tracks — targeting the quality of the original Secret of Monkey Island soundtrack.
The result, according to Willison’s account, was “surprisingly good.” Claude did not just generate a description of music. It designed a lightweight text-based music notation format, wrote code to interpret that format and play it audibly in the browser, and then composed original tracks in that format — all in one go. Willison notes that it “leaned a lot harder into the Monkey Island theme” than he intended, meaning the model ran with the creative brief more literally than expected, but the output was musically coherent enough to make him sit up and take notice.
What Is Actually Happening Here
To understand why this is interesting, it helps to break down the three things Claude did simultaneously in this experiment.
The whole chain — format design, tooling, composition — happened in response to one prompt.
Why Willison Is Treating This as a Possible Capability Shift
Willison is careful not to overclaim. He explicitly says he would “need some careful experiments with other recent and not-so-recent models to confirm if this is new or if they’ve been able to do this for a while.” He is not declaring a breakthrough. What he is doing is flagging a pattern.
He draws a comparison to 3D graphics — another domain where text-based AI models appeared to develop competent generative ability relatively recently. His question is whether music composition, specifically the ability to produce something structurally coherent and aesthetically recognisable, is a capability that has emerged in large language models over the past few months, or whether it has been latent for longer and he is only noticing it now.
This distinction matters for anyone tracking what these models can actually do versus what they have always been able to do but nobody thought to ask. The Scrimshaw Jukebox experiment is, at its core, a probe: what happens when you ask Claude to invent the rules of a creative system and then play by those rules at the same time?
What This Looks Like for a Non-Technical User in India
Consider a small indie game studio in Pune — perhaps two or three people working on a mobile puzzle game aimed at the Indian market. They have a designer and a developer, but no composer. Hiring a professional composer or licensing a royalty-free soundtrack library can cost anywhere from ₹15,000 to several lakh rupees depending on scope and exclusivity.
The Scrimshaw Jukebox experiment suggests a workflow that did not exist in a practical form until very recently: describe the mood, era, and style of music you want, ask Claude to design a playable format for it, and receive a functional, browser-playable musical prototype in a single session. That prototype may not replace a professional composer for a AAA title, but for an indie team testing whether chiptune-style background music fits their game’s atmosphere before they commit budget, it is a meaningful capability.
The same logic applies to a content creator in Hyderabad producing short explainer videos for YouTube, or a small advertising agency in Chennai that needs placeholder music for a client presentation before the actual audio production budget is approved. The ability to rapidly prototype audio — not just describe it — changes what is possible at the concept stage.
The Honest Limitations You Should Know
Willison’s post is a brief, honest account of a single experiment, and it is important to read it that way. Several things remain genuinely unclear or constrained.
What to Watch For Next
Willison himself flags the most important next step: comparative testing. Does this capability exist in earlier Claude versions? In models from other providers? In smaller, locally-runnable models? Until those comparisons are done systematically, it is hard to know whether this is an Opus 5.5 specific capability, a Claude-family capability, or something that most frontier models can now do.
From a practical standpoint, watch for whether Anthropic or third-party developers build more structured tools around this kind of generative audio workflow — tools that add audio export, format standardisation, or DAW integration on top of the raw model capability that Scrimshaw Jukebox demonstrates.
If you want to explore the experiment yourself, the most direct starting point is reading Willison’s original post at simonwillison.net, understanding the exact prompt he used, and attempting a version of it with whatever Claude tier you have access to. Pay attention to whether the output is actually playable, whether it sounds musically structured to your ear, and whether varying the stylistic reference changes the output in predictable ways. That kind of careful, iterative probing is exactly what Willison is modelling — and it is a more useful approach than assuming the result generalises immediately.
