Diorama Journey

Diorama Journey

Tell a story by following something through a small, connected 3D world. The recurring element carries meaning between chapters; its actions and changing surroundings explain the subject. Deliver a narrated film and presentations from the same composition.

Use this for miniature 3D storytelling. Ordinary screen recordings, flat motion graphics and general slide authoring belong to other workflows. The input can be an outline, report, lesson, process, repository or user-provided material. No specific organization, product, account or publishing service is required.

Shape the story

Identify the audience, takeaway, language, target duration and requested deliverables from the user's material. Resolve routine choices from context; ask only for missing decisions that materially change the work. The starter uses 1920 × 1080, 30 fps and English narration. Change the voice and text for the requested language. Other canvas sizes require adapting the composition and checking the actual layout.

Read story and art direction. Choose one recurring character, object or signal and explain what it represents. Give each chapter one claim, an action and a visible consequence. Preserve that same element through the transitions, and make the ending resolve the opening.

For factual content, record the source behind claims in each scene's evidence array. Separate the user's confirmed results from proposals. A fictional example can have empty evidence; never invent measured outcomes to fit the scene.

The included little-light story is a runnable example. Replace its writing and landmark actions for the new subject. A JSON action describes intent; implement the geometry and motion in src/art/World.tsx. Do not assume editing prose has changed the animation. The provided gate, stairs, workshop, bridge and garden can be reordered or repeated; add code for a genuinely different action.

Create an independent production

Use the actual absolute path of this skill in place of SKILL_DIR:

python3 SKILL_DIR/scripts/doctor.py
python3 SKILL_DIR/scripts/create_project.py my-story --output ./productions

The helper creates productions/my-story/UTC_TIMESTAMP/, copies the runnable starter and creates run.json. Existing runs are never overwritten. Work in the new run, not in the installed skill. Keep project-specific source material and publication settings there.

Update src/story.json, scene code, HTML metadata and the production brief. Keep the narration, displayed copy and physical actions consistent. No fixed scene count is required; every story scene gets timing, a clip and a slide.

Produce and review

Read production for the environment, commands and rendering invariants. Read deliverables and preservation before rendering the final set.

1. Install the pinned dependencies. Checkpoint before replacing a meaningful draft. Generate local font subsets for new glyphs, narration and the music bed. Narration determines the timeline and sentence subtitles.
2. Run type checking, the build and asset verification. Render every chapter's keyframe, a clean cover and a transition sample. Inspect all stills and the moving sample before the full film. Check the motif's continuity, readable text, clear actions and the camera path.
3. Render the mastered MP4, clean slides and chapter clips. Export PPTX, PDF, ODP and full speaker notes. Run verification; retry only failed stages.
4. Inspect the encoded video, listen to the narration and music, and review the exported slides. Check the viewing site's playback, chapter seeking, downloads and narrow layout. Report checks that could not run precisely.
5. Build the viewing site after the deliverables exist and archive the finished production. Provide paths or links to the output set and its verification record. Keep visual or listening feedback as human review when applicable; automated checks do not establish those qualities.

The full command is npm run render: video, slides, then clips. Use npm run slides to regenerate only presentations and npm run clips for slices. The default delivery includes all of them; honor an explicitly smaller scope.

Important mechanics

• Derive motion from the requested frame. Keep scene-local motion based on that scene's own start, including incoming and outgoing scenes. Unseeded randomness, wall-clock animation and advancing render loops break seeking and exports.
• Keep a full-size ThreeCanvas. Move the camera or view offset, not a small canvas that clips the world during travel. Inspect WebGL in the actual render.
• Pass identical inputProps to composition selection and rendering. Clean covers and slides use captions:false; otherwise resolved props can put subtitles back into those outputs.
• Keep narration MP3, boundaries and cache signatures together. After speech changes, regenerate timing and music before images, video, clips and slides. Preserve valid cached audio if an online speech request fails.
• Chapter clips use exact frame boundaries and are re-encoded for accurate cuts. Their sidecar subtitles are clipped and rebased to chapter time.
• PPTX and ODP contain full-frame images with editable native speaker notes. Their 3D objects and headlines are not independent editable slide shapes.
• Preserve original inputs, licenses, source snapshots, render attempts and verification records. The archive excludes rebuildable bundles and caches. Keep large production archives outside the skill's source repository.