SuperAPI AI Directory

narova-3d-production

Design & creation

View repository

Updated 10-16 03:13

About this skill

Narova 3D Production

Add a small expert lens to authored 3D work. Keep the narova skill in charge of project authoring, assets, canonical time, rendering, review, release, and delivery. This companion can optionally bake bounded local rigid-body motion; it adds no core dependency, renderer, provider, hosted service, universal scene graph, or second production lifecycle.

Direct lightly

  1. Give 3D a job. State what authored 3D lets the viewer perceive that a simpler or different medium would not. If it adds no material value, use the better Narova path and stop using this companion.
  2. Choose what the audience will receive. Decide what the viewer should notice, understand, or feel, then choose intentional final forms for the principal subjects and world. A primitive, model, captured or generated element, and mixed composite are all valid. Do not silently turn a blockout or cheapest available shape into the final representation.
  3. Route the material risk. Choose at most the one or two areas most likely to weaken the idea and read only their references:
    • read subjects and assets when final representation, acquisition/generation, model readiness, rigging, animation, rights, or a missing production capability matters;
    • read scene direction when space, appearance, staging, camera, visibility, interaction, or compositing is uncertain; and
    • read physical reasoning when support, locomotion, collision, constraints, gravity, coupled motion, or scientific accuracy is a material risk; and
    • read inspection when the visual premise, movement, structured state, evidence limits, or a stop decision matters. A straightforward scene may need none. Do not perform a full-manual pass.
  4. Prove before scaling. For risky work, show one delivery-scale visual target without its rationale. If motion or interaction carries the idea, next show a short sample using that accepted representation. Expand only a premise that reads from the artifact itself.
  5. Finish through core Narova. Use available specialist capabilities only when the chosen form needs them, then hand assets, deterministic motion, rendering, review, release, and delivery to the ordinary Narova lifecycle.

Bake rigid-body motion only when it helps

Prefer authored motion when timing, exaggeration, or expressive control carries the idea. When gravity, support, collision, stacking, or a rigid constraint carries the idea, install this companion's local tool dependency and bake an explicit project-local recipe:

cd skills/narova-3d-production && npm ci --ignore-scripts
node tools/bake-rigid-body.js /absolute/project physics.json physics-bake.json

Sample the bake with tools/sample-bake.js at canonical local scene time; do not advance a solver while rendering frames. The recipe and bake are execution inputs, never sceneState evidence. The built-in baker is limited to bounded rigid bodies and simple constraints. Route cloth, fluids, fracture, robotics, or engineering/scientific verification to an appropriate specialist.

Use an optional DCC target only when the chosen form needs it

For Blender-specific work, read DCC environment and operations. Assess the installed target and workload; availability alone is not suitability. Never install or transmit implicitly, pretend unsupported breadth is available, or silently degrade a required subject, relationship, behavior, or deliverable.

Inspect the exact shot source that will render. Sparse frames expose shot-local program facts. Explicit proof requests can expose decoded-pixel distributions and coarse requested-object camera projection. Narova does not choose the frames, set target bands, infer what the audience sees, declare completeness or smoothness, or recommend a creative correction.

A committed DCC result enters ordinary Narova as an authored local media asset; core still owns time, audio, composition, proof, release, and delivery. No house template, style, camera, realism, or prior-project memory is imposed.

Understand and encode a committed frame sequence

After an agent commits frames, the companion can encode a bounded MP4. It explains observable effects without ranking or selecting; the agent supplies every creative encoding value.

Interpolation modes and their observable consequences:

Mode What FFmpeg does Observable effect
hold Holds, drops, or duplicates source frames to match output FPS No motion-estimated frames; motion may appear stepped when rates differ
blend Creates intermediate frames by blending adjacent source frames Motion may appear smoother; blur or ghosting can appear
motion-compensated Synthesizes intermediate frames from estimated motion Motion may appear smoother; distortion can appear at occlusions, fast motion, or scene changes
node tools/frame-sequence-to-mp4.js /absolute/project encode-request.json encode-receipt.json

Read frame encoding reference for the complete request, neutral consequence descriptions, validation, and receipt format.

Keep confidence honest

Connect decisions to what was rendered or otherwise inspected. Keep factual checks separate from perceptual judgment. Do not give the perceptual reviewer the author's plan, symbol mapping, or run report. Treat an automated visual critic as advice until it has agreed with independent owner judgments on this kind of work. Render success does not prove recognition, collision, physical simulation, target visibility, taste, or hidden scene state. Non-black frames, motion, clearance, sightlines, render-enabled state, camera deltas, projected bounds, decoded-pixel distributions, and sparse samples prove only their named properties; inspection owns broader claims.

If the chosen form needs sourcing, generation, refinement, rigging, animation, simulation, or inspection that no available capability supplies, change the form with the user, name the gap, or stop. Never pretend this skill's prose performed that operation.

Preserve creative range

Do not default the work to a palette, realism level, low-poly language, material system, lighting rig, camera move, shot list, density, pacing, duration, narration, caption style, transition, platform, or aspect ratio. Minimal and rich scenes, graphic abstraction and articulated realism are all valid when intentional. Use hard constraints only where user intent, compatibility, evidence, or visible correctness would otherwise break.

Do not persist or retrieve another project's concept, prompt, subjects, assets, world, palette, camera, style, or owner disposition. Let agents grow through current-task reasoning, executable consequences, and bounded proofs without anchoring the next work to a sibling solution.