Moûsa

Design for an audience from what it actually loves. Moûsa is a plugin for design agents in Claude Code, Codex and Kiro. Before the agent designs anything, it freezes its own guess about the audience. Moûsa then measures the audience with Qloo and binds the design to what the guess missed. Every design ends with a report that sets the guess beside the measurement and shows where each choice came from.

An entry to the Qloo Agentic Hackathon.

For the jury: the video, 1 minute 47 seconds · the site, which tells one pair from the request to the record, with both games at play · Moûsa in one minute · the showcase, a page that explains Moûsa, designed by Moûsa · the benchmark · the live try, open to anyone with no code, up to 100 first steps and 10 design pairs a day.

The problem

Teams now build interfaces with AI coding agents, and those agents design for audiences they have never met. Ask one for a record shop's website "for people who love Paterson and Nujabes", and it designs from its own idea of those people: in a live run, Kiro's sealed guess named one of the 69 works Qloo measured for that audience. The site looks plausible and generic, because that idea was never tested.

How large the gap is, measured. In a pre-registered benchmark, three agents froze their own predictions for 20 audiences. Every prediction was committed to this repository before the benchmark asked Qloo: GitHub's push log shows them arriving at 14:14 UTC on 23 October, and the analysis is stamped 14:21. Pooled, the guesses missed 94.0% of the works Qloo then measured. The guesses were shorter lists than Qloo's (Claude Code named 571 works against the 1,395 measured), which caps that share, so the table also counts the other way: of the works each agent guessed, how many Qloo measured. Each agent's guesses named this share:

Agent	Known share (95% interval)	Guessed works Qloo measured
Claude Code (claude-opus-5-5)	6.8% (5.2% to 8.5%)	16.1%
Kiro (claude-opus-5.5)	6.6% (5.2% to 8.0%)	15.8%
Codex (gpt-6.1-sol, high effort)	4.7% (3.2% to 6.1%)	13.9%

Both pre-registered thresholds held for every agent: under 15% of the measured works named, and under 25% of the guesses measured. The agents knew music and TV best (about 10%) and films, brands and books least (2% to 6%). Claude Code and Kiro run the same model and differ by 0.2 points, which suggests the blind spot belongs to the model rather than the tool. Results, predictions.

Designers, brand teams and game studios who work with agents need the agent to know its audience beyond the obvious, and to show its evidence. Moûsa does this inside the agent a team already uses, on the project it already has: a website, an app's design tokens, a game character.

How it works

1. The agent guesses first. Moûsa picks the taste questions for the medium: which Qloo domains to measure (films, music, books, TV, brands, places, games) and which tags. Before any Qloo data, the agent writes its own prediction of what the audience loves, and the design it would make from that alone. The prediction is frozen and sealed with its SHA-256.
2. Qloo measures the audience. Moûsa asks Qloo for insights in each domain and for the audience's style and aesthetic tags. It then groups the measured works' own tags into shared qualities, ranked by how distinctive each is for this audience against 28 others. A session sends at most 24 requests, and the answers are cached for a week.
3. The design must rest on what the guess missed. The agent reads the measurement, names how the audience's works look as a genre (broadcast score bugs, mixtape covers, hand-painted market signs), and writes a Design DNA: concrete devices for palette, type, motion, illustration and voice, each anchored to what Qloo measured, with the palette or type drawn through one of those conventions. Moûsa checks every device in the project's own design system and flags the looks models default to (a cream page, a stock display serif, monospace labels, pill buttons, a near-black page lit by one amber lamp and the like; at most two may stay). The agent reviews screenshots of what it built, names its three most generic choices and asks whether a fan of the audience's best-loved works would take the page for theirs.
4. The record shows it. The Taste Provenance report sets the guess beside the measurement, marks each measured name as new or predicted, and follows it to each device and to the screenshots reviewed.

So Qloo's contribution is visible as a difference: what was measured, against what the agent would have assumed.

The evidence

Each kind shows something different, and none shows everything. The jury page sets them side by side, each with what it does not show.

1. The blind spot. The benchmark above. Separately, in the 24 recorded sessions in the gallery, the guesses named 62 of the 1,618 works Qloo measured, and 62 of the 597 works they guessed were measured. The sessions share a few audiences, so the counts overlap. Anyone can recount that from the release's gallery bundle with node scripts/blind-spot-stats.mjs <folder>.
2. The mechanism. Each comparison builds two designs from one frozen prediction: one with Qloo's measurement, one from the prediction alone.

Comparison	Agent	Distance between the two designs	Devices resting on what the prediction missed
Moûsa, explained to the jury (the showcase)	Claude Code	0.63	10 of 10
Moûsa, explained to the jury, first build	Claude Code	0.60	8 of 8
The Lamplight, a live music venue	Claude Code	0.62	7 of 7
Second Drink, a stand-up night	Claude Code	0.69	8 of 8
Thistledown, a browser RPG from one sentence, for Stardew Valley fans	Claude Code	0.62	9 of 9
Saltmarsh, a browser RPG for Wes Anderson and Fleetwood Mac fans, no game named	Claude Code	0.57	9 of 9
Opening Night, a browser game for the jury's own audience	Claude Code	0.66	9 of 10
Night Shift Run Club, a phone app	Claude Code	0.63	10 of 10
Ilo, a game character	Claude Code	0.53	9 of 9
Ilo, a game character	Codex	0.57	7 of 7
Ilo, with painted portraits	Codex	0.58	5 of 5
A record shop for Paterson and Nujabes, live try, 24 October	Kiro	0.65	7 of 7
A record shop for Paterson and Nujabes, live try, 14 October	Kiro	0.60	7 of 7
A record shop for Taylor Swift and Gilmore Girls, live try	Kiro	0.68	7 of 8
A record shop for Bad Bunny and Narcos, live try	Kiro	0.65	7 of 7
A record shop for Elden Ring and Berserk, live try	Kiro	0.74	7 of 7

The distance is a heuristic over palette, type, page structure and CSS, or for a game the parts and words its code names, from 0 (alike) to 1 (unrelated). It does not rate quality. The protocol requires the names the prediction missed to change the Qloo design's palette, type and structure; the last column counts the devices that rest on such a name.

How it plays, not only how it looks. Four of these pairs began from one sentence and an empty folder. Their arms differ in what they do. For Stardew Valley fans the guess built farming, fishing and friendship; the measurement built herding from Slime Rancher, tape-recorded interviews from a Letterman interview show and 63 Up, and a band rehearsal from Modern Baseball. For the jury's audience the guess stitched the five names into an at-bat, a roast, a pitch room and an encore; the measurement made one stand-up set timed like a competitive game. See the verification record. 3. Beyond the model's own variance. In repeated pairs of one website brief (12 generations), the blind sketches of both arms showed no detectable difference (p = 0.81). After Qloo, typography, structure, palette and overall distance separated (exact permutation test, p = 0.005 each). For game characters, the finished characters separated (p = 0.0079) and the blind sketches did not. The protocol asks the Qloo arm to move from its sketch, so this shows that the move is more than the model's own variance, not that the moved design is better. See the development notes. 4. Preference. The blind vote shows a visitor both designs without saying which one used Qloo, and asks which the audience would rather have. By 27 October no votes had been cast; the results page counts each one.

The live try runs Moûsa in Kiro on the server for an audience the juror names. Step one, under a minute, freezes Kiro's prediction and measures the audience. Step two, about five to ten minutes, builds both designs.

Install the plugin

Each person uses their own Qloo API key. It goes into the host's plugin settings or onto Moûsa's local key page (mousa_connect), never into the chat.

# Claude Code
/plugin marketplace add tritonsan/mousa
/plugin install mousa@mousa

# Codex
codex plugin marketplace add tritonsan/mousa
codex plugin add mousa@mousa

# Kiro, V3 engine, from a clone of this repository
kiro-cli powers install plugins/kiro

The plugin brings 13 tools, four skills (taste-brief, design, compare and connect) and a library of 55 open-licence font families. In Claude Code and Codex, a write gate also holds the design's files until its Design DNA is valid. See the plugin README.

How it is built

• plugin/server is the engine: one MCP server in Node 22, with no dependencies. It plans the taste questions, calls Qloo within a budget and a cache, freezes the prediction and writes the Taste Brief. It validates the Design DNA and checks it through the project's CSS custom properties, token files or Tailwind theme. It captures pages in a headless Chromium that can reach only the page it captures, and it writes the reports.
• plugins/ holds the Claude Code, Codex and Kiro packages, generated byte for byte from plugin/.
• src/ is the hosted site: the gallery, the blind vote with sealed ballots, and the live try. The try runs each Kiro session with an agent file of its own: Moûsa's tools, no shell, and write access to the starter's design files only.
• deploy/ holds the Docker image, the Compose file and one CloudFormation stack for AWS. The image runs with a read-only root and no capabilities, under a seccomp profile that keeps Chromium's sandbox.
• tests/ holds 559 tests, with no network or model call, run on Ubuntu and Windows in CI.

Run it locally

Node.js 22 or later. The live try also needs the Kiro CLI, signed in or with a KIRO_API_KEY.

cp .env.example .env    # then set QLOO_API_KEY
npm test
npm start               # http://127.0.0.1:8787

Deploy

The server is one CloudFormation stack: an instance behind two CloudFront names, the keys in Secrets Manager and a daily snapshot. See AWS deployment.

Limits

• A distance shows that two designs differ, not that one is better. The repeated-pair studies each used one brief and one model.
• The benchmark measures what each agent assumes when Moûsa asks it, with one prompt and one run per audience. Qloo's measurement is aggregate affinity data, not ground truth about any person.
• The vote is self-selected and small.
• The hosted server has no image generator, so designs made there draw their imagery in SVG and CSS. Recorded sessions made with Codex's image tool say so.
• Qloo's answers stay on the machine that asked for them, as Qloo's developer guide asks of a public repository. The site and this repository show the names Qloo measured, never its ids, scores or descriptions: the two recorded runs kept here carry stand-in ids and no scores, through the whole history, and a test keeps it so. Each design choice is the agent's interpretation of the names shown.

More

• Verification record: every live run, dated, with what it did not show.
• The benchmark: its protocol and amendment, the sealed predictions and the results.
• The submission text: the Devpost entry, with the path from a request to its result.
• Qloo integration and creative policy.
• Qloo findings: what the live API taught us, what the measurements showed, and what we would ask of the API.
• Development notes: the earlier, detailed README, with the protocol's history and the Studio.

License

MIT. The fonts, photographs, music and seccomp profile carry their own licences; see the third-party notices.