| name | palace-visualizer |
| description | Render a memory palace as a navigable artifact — a self-contained HTML viewer plus image-generation prompts for each locus. Use when the user says "render my palace", "visualize the palace", "make a picture of my memory palace", "build out the palace", or wants a keepable visual. Reads palace.json, emits a standalone index.html (no build step) and per-locus image prompts; tolerates both memory-palace and blessing schemas. |
palace-visualizer
The render step — build the palace out into something you can see, walk, and keep. The agent
helps make the imagined place concrete.
When to use this
A palace exists in palace.json (or a Blessing rooms.json) and the user wants a visual artifact:
a navigable HTML scene, and/or image prompts to generate art for each locus.
Instructions
-
Read the source and branch on type. "memory-palace" → render the nested
room → station → locus hierarchy. "blessing" (or a rooms.json) → render the flat rooms. The
shared surface/accent vocabulary means one renderer serves both.
-
Emit a standalone index.html. Start from assets/palace-viewer/index.html and inline the
palace data so the file works offline with no build step and no fetch. One <canvas>, Three.js
via CDN importmap + ESM, with a pure-CSS fallback if 3D is declined.
-
Hold the craft discipline (identical to palace-build, so both palaces look like one family):
- One warm key light (
#ffe5cc) + one cool rim (#7da3ff) + low ambient. Not five lights.
- Dark background (
#09090b). Selective bloom on emissive accents only — never global.
- Glass that stays legible (thickness ~0.25, low chromatic aberration). Labels must read.
- Damp camera motion; bind reveals to input. Respect
prefers-reduced-motion (disable idle drift).
- Each locus is a named plinth/orb; its
image.text (or room truth) is legible on hover/select.
- Map each room's
surface to assets/textures/<surface>.webp, falling back to flat accent
colour if the texture is missing — so a lone copied index.html still renders.
-
Generate per-locus image prompts. For each locus, derive an image-gen prompt from
image.text (or write one if absent) and store it in image.prompt. Prompt template:
"<image.text>, dramatic single key light, dark background, photographic, highly detailed,
no text". These can be rendered via the ~~image-generation connector (e.g. Higgsfield) and the
result path written back to image.asset.
-
Keep palace.json the source of truth. The viewer is a projection; never edit facts in the
HTML. If the palace changes, re-render.
Refusals
- Never overcrowd the scene — if a room holds many loci, group or summarise rather than spawn clutter.
- Never bake external data fetches into the standalone file; inline the data.
- Don't invent rooms/loci not present in the source; render what is there.
References
assets/palace-viewer/index.html — the standalone viewer/builder this skill ships and inlines.assets/textures/ — the six surface textures.spec/MEMORY-PALACE-METHOD.md §5; Blessing SPEC §5 for the rooms.json branch.
Built on SIP · mind-palace-agent-skills · Memory Palace Method v0.1
