Catalogue essay · FABLE / 175

How this was made

A practical account of seven wings, the people-and-model governance, the audit harness, and the visual production system behind one hundred seventy-five autonomous websites.

The commission

FABLE / 175 began as a deliberately unreasonable assignment and kept expanding: create one hundred seventy-five fundamentally different websites, each with its own fictional institution, object, product, place, or machine. No shared template. No common palette. No house layout disguised in 175 costumes. Every site needed its own palette, type system, signature technique, interaction logic, and an unforgettable moment. Each one also needed a local /guide explaining how it was made. The work was built autonomously, then forced through three documented iteration passes so the final gallery would feel like a set of rooms, not a set of drafts.

The governance model

Claude Fable 5 acted as planner, architect, and art director. Fable wrote all 175 concept briefs: the palette, type, signature technique, emotional posture, and the exact thing that had to make the page memorable. Codex (gpt-5.6-sol, high reasoning effort) acted as executor: reading the brief, building the site, auditing it, and carrying fixes through until the page matched the direction. Opus stayed on standby as a coding fallback, but the main pipeline remained Fable direction into Codex execution.

Each site followed the same governance rhythm. Pass 1 was build plus self-critique: make the thing real, then inspect it without politeness. Pass 2 was a merciless external critique: composition, responsiveness, load, overflow, dead corners, broken illusion, weak typography. Pass 3 was art-director elevation and certification: not merely clean, but worthy. After that came the Fable taste gate on final screenshots. That gate was allowed to say no. The HELIOS-style visual regression caught in Wing I never recurred. Wing II produced a different save: CIRCUITRY's Chromium audit process crashed under load, so the architect reran the evidence capture instead of mistaking a broken audit for a broken room.

Wing II

The second commission arrived with a simple escalation: make twenty-five more, and make them even more stunning. Wing II was not a batch of extensions to the first set. It added a from-scratch build phase in which every room began from an empty directory and a fresh brief, then earned its way through the same three-pass discipline. The governance did not change with the ambition. Fable still wrote the world, set the visual gates, read the screenshots, and decided whether a room had become itself; Codex still executed the build, ran the audits, and carried corrections through to clean evidence. Repetition belonged to the process, never to the sites.

The asset vocabulary widened as well. Generated stills and Kling motion remained part of the material language: rain moved outside the NIGHTHAWK diner, the ABSINTHE muse breathed inside the green hour, and the ARCANA box became a charged object rather than a static prop. BLUENOTE added a new modality—generated audio—with a seed_audio jazz trio composed for the basement club. The sound was treated like every other asset: local, staged by the interface, and subordinate to the room's atmosphere rather than added as spectacle.

Scale exposed different operational failures. Chromium could SIGTRAP under the load of dense visual audits; when CIRCUITRY's capture failed, the architect reran the audit and judged the replacement evidence instead of lowering the gate. OpenAI 429 rate limits interrupted long execution rounds, so the work continued through session-resume rounds rather than being restarted or waved through. Those incidents clarified the system's real strength: not uninterrupted automation, but recoverable craft. Wing II raised the bar because the same authorship model could survive new media, heavier pages, and broken runs without relaxing what “finished” meant.

Wings III & IV

The second doubling added fifty new briefs at once, but scale was never allowed to become sameness. Wings III and IV were commissioned against a zero-aesthetic-overlap rule: each room needed a new visual grammar, material logic, and memorable interaction. The territory moved deliberately through craft, science, and performance. Letterpress sat beside glassblowing; falconry physics beside ice cores; foley sound beside rooftop kite duels; English change-ringing beside diffusion-limited aggregation frost. The subjects were not decorative themes. Their real procedures shaped the interface, from mirror-set type and molten gathers to bell methods, tide windows, flight forces, and crystal branching.

The asset vocabulary widened without replacing code as the staging intelligence. DARKROOM develops real Higgsfield prints under red light instead of showing anonymous placeholders. NICKELODEON hand-cranks a generated film loop, with Kling flicker bound to the visitor’s wrist. TEAGARDEN carries a mist loop through the terraces, while CAROUSEL turns a seed_audio calliope waltz into part of the ride’s machinery. These assets behave like physical stock inside their rooms: exposed, cranked, weathered, projected, or played. They are locally held evidence that the fictional institution has a material culture, not interchangeable decoration pasted behind the layout.

The production line still held at roughly five finished sites per hour. The only automation needed to keep the queue moving was self-healing retry logic for 429 rate limits; every visual judgment still passed through responsive screenshots, console evidence, written critique, and the final taste gate. The last dependency changed the order of work. Thumbnails had to be made before the finale because CENTENNIAL, the hundredth room, uses the other ninety-nine rooms as textures in its orbiting atlas. The catalogue therefore became source material for its own conclusion: ninety-nine independent surfaces completed first, then gathered into one navigable constellation whose doors still open back into the worlds that made it.

Wing V — The War Wing

The fifth wing changes register without changing the exhibition’s discipline. It is loud, dark, and cinematic: heroes arrive as movie frames, letterbox chrome compresses the view, and film-gate numbers, reticles, timecodes, and archive annotations make each page feel found rather than merely displayed. The color grades lean toward teal and amber, rain-black steel, desert brownout, and the acid phosphor of night-vision goggles. Even when the hero is a still, the frame behaves like a shot. Grain crawls, smoke drifts, rain tracks the glass, snow crosses the lens, and heat loosens the horizon.

That register came with firm guardrails. Every room declares itself as fiction, history, or memorial. Fiction uses invented units, ships, call signs, stations, and missions. History looks to Sekigahara in 1600, the age of sail, Roman campaign life, siege craft, and preserved twentieth-century machinery. Memorial rooms hold the Cenotaph vigil, empty dispersals, and a war photographer’s withheld archive. There is no operational instruction, no contemporary conflict presented for imitation, and no real war glorified. Violence remains off-screen. The work chooses anticipation, consequence, restraint, and remembrance over gore.

Audio follows the same ethic of consent. WebAudio synthesizes atmosphere behind explicit switches that default off; two rooms are deliberately silent, and CENOTAPH is still-only by design, holding its image rather than converting the vigil into spectacle. The image budget shaped the rest of the wing. All twenty-five rooms received GPT Image 2 cinematic stills, but Higgsfield credits ran out midway through production. Only seven Kling loops exist: GHOSTDIVISION, FLIGHTDECK, SANDSTORM, SEKIGAHARA, BOMBERGROUP, WHITEOUT, and SILENTRUNNING. The other eighteen heroes stage stills like film frames, using code-driven rain, shimmer, grain, smoke, parallax, and interface motion to imply a camera without pretending a photograph is video. A shortage became a compositional rule, then a technique: pressure without spectacle.

Wings VI & VII

Wings VI and VII arrived together, but they speak in opposite registers. Velocity is bright, engineered, and kinetic: signal orange across a lightened field, speed expressed through motion design rather than borrowed photography. Aero-drag scroll makes resistance legible; g-force typography stretches under load; telemetry chrome turns timing, slip, temperature, and pressure into the page’s nervous system. Its teams, championships, vehicles, institutions, and machines are all fictional, allowing the visual codes of motorsport, flight, rail, cycling, sailing, and human performance to be studied without pretending to document real competitors.

Nocturne is the deliberate anti-War Wing. It is dark but soft, interested in wonder rather than menace. Light itself becomes the design material: a Bortle slider gives the Milky Way back to a polluted sky; a headlamp cursor reveals a cave only where attention falls; fireflies learn to pulse together; gas lamps are kindled one by one. Other pages accumulate starlight as the visitor waits or shed power until a town discovers the people still awake inside it. Indigo, amber, eyeshine, bioluminescence, freezer glow, polar blue hour, and the thinning light before dawn turn darkness into an active, generous medium.

The production constraint became the clearest flex in the exhibition. These fifty rooms used zero generated assets. Image credits had run out, so every visual in both wings is code: canvas fields, SVG instruments, CSS atmospheres, procedural drawing, typographic systems, and WebGL wind or flow. The absence is stated proudly because it sharpened the briefs. Nothing could hide behind a hero render. Every memorable image had to be built as behavior, and every behavior had to belong to its subject.

The exhibition also began documenting its own recreation. Every one of the 175 rooms now ships a Prompt link in its footer. It opens the actual art-direction brief that produced the room, copyable and ready to paste into any AI coding agent. The catalogue is therefore not only evidence of the finished work; it carries the instructions for making its worlds again. The final expansion took 150 Codex runs. One failure was recovered when a resumed session lost Chromium access mid-audit: the architect ran the audits directly, preserved the evidence, and re-dispatched the work fresh. Across all fifty rooms, zero taste-gate rejections were recorded.

The quality harness

The harness was intentionally blunt. Every pass used Playwright screenshots at 390, 834, and 1440 pixels wide, covering top, middle, bottom, and full-page states. Console errors, page errors, failed requests, and horizontal overflow were treated as hard evidence. External requests were forbidden; every font, image, video, and script had to resolve locally. Horizontal overflow was an immediate fail because a page that cannot respect the viewport cannot be called finished.

Exit codes were helpful but not trusted as truth. The evidence queue was validated by files: each site needed its PASSES.md record, screenshot labels for every pass, and CLEAN audit logs before it could leave the queue. Long-running work used codex exec resume so the process could continue without pretending memory was a plan. The point was not bureaucracy. It was to make taste accountable to screenshots.

The visuals

In Wing I, seventeen sites are 100% code. They use GLSL and WebGL-style shader work, Three.js terrain and particle renderers, L-system SVG, canvas particle systems, SVG filters, Web Audio, procedural typography, and CSS choreography. HELIOS, TERRANE, FOLIUM, KINETIKA, ABYSSAL, UKIYO, CHROMA, CARTOGRAPH, RESONANCE, VERSE, BRUTA, STRATUS, HEXADOME, TERMINAL NINE, KANTON, THE PAPER FOX, and SINGULARITY all stand on code as their visual engine.

The other eight Wing I sites blend code with AI-generated key assets: NOIRE, ESSENCE, DYNASTY, LUMEN, MERIDIAN, RELIQUARY, VELLUM, and APOTHECARIUM. The stills were generated with GPT Image 2 through Higgsfield, then animated into five-second ambient loops with Kling 3.0 using image-to-video from our own stills. Those assets are stored locally and treated as source material, not remote decoration. The code still does the staging: motion, masks, scroll, type, interaction, and the final page rhythm.

Reproduce this

Start with brutal-specific briefs. Name the world, the audience, the palette, the type, the signature technique, and the moment the viewer should remember after closing the tab. Forbid the generic before it appears: no default fonts, no default palettes, no interchangeable card stacks, no decorative gradients pretending to be art direction. Make the brief narrow enough that a bland result is obviously disobedient.

Then force three documented passes. Capture screenshots at phone, tablet, and desktop. Read the console. Reject horizontal overflow. Write down what changed and why. Put a taste gate at the end that is allowed to say no, even when the page technically works. Especially then. The durable lesson of FABLE / 175 is that autonomy only becomes craft when it is paired with evidence, refusal, and a final human-scale eye for whether the thing has become itself.