narova-3d-production · diff

v0.3.0 to v0.6.0

46 added, 7 removed. Audit A to A.

---

name: narova-3d-production

description: >

Use this optional Narova companion when creating, directing, or repairing an

authored 3D video whose principal subjects, world, asset path, spatial design,

materials, lighting, camera, interaction, or inspection need specialist

judgment. Use for abstract or realistic 3D, recognizable or articulated

subjects, layered environments, mechanisms, camera-led scenes, and mixed

2D/3D where authored 3D has an intentional role. Do not use it merely because

a request says animation, cinematic, explainer, product video, Three.js, or

render; ordinary 2D motion, browser walkthroughs, AI-generated clips, and

direct render commands remain core Narova work.

license: Apache-2.0

metadata:

author: ammar-hasan

- version: "0.3.0"

- checksum: 1d07fccc7808733087636d7eae0974688ac619af898a368bd56b08ca361a4586

+ version: "0.6.0"

+ checksum: cad4c393fc11fb51c22fc16a70529f8f8a79fe094a855456098d3f627db8925a

---

# 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](references/subjects-and-assets.md) when final

representation, acquisition/generation, model readiness, rigging,

animation, rights, or a missing production capability matters;

- read [scene direction](references/scene-direction.md) when space,

appearance, staging, camera, visibility, interaction, or compositing is

uncertain; and

- read [physical reasoning](references/physical-reasoning.md) when support,

locomotion, collision, constraints, gravity, coupled motion, or scientific

accuracy is a material risk; and

- read [inspection](references/inspection.md) 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:

```sh

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](references/dcc-environment.md).

+ 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 |

+ ```sh

+ node tools/frame-sequence-to-mp4.js /absolute/project encode-request.json encode-receipt.json

+ ```

+ Read [frame encoding reference](references/frame-encoding.md) 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.

+ 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](references/inspection.md) 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.

+ 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.