Skip to content
Back to skills

Cinematic Pack Authoring

ASecurity

Author a Slopcamera spatial recipe pack — a bounded, inert, content-addressed JSON document that names an admitted scene, one direction document, gallery axes, preview render requests, declared effects, and a temporal audit — then plan or run the built-in cinematic-world workflow. Use when a user asks an agent to direct a 3D scene, iterate camera/look/sequence variants, or review compiled direction before explicit selection.

  • 7 stars
  • 0 votes
  • 0 copies
  • 2 views
  • Added September 29, 2026
ai-agentsperformance

Security analysis

A100/100

Pro scans all 3 files and shows the line behind each finding

Scanned September 29, 2026

npx -y skills add hraness/slopcamera --skill cinematic-pack-authoring --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Cinematic Pack Authoring?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Cinematic Pack Authoring
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/hraness-cinematic-pack-authoring/badge)](https://www.skillsdirectory.com/skills/hraness-cinematic-pack-authoring)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: cinematic-pack-authoring
description: Author a Slopcamera spatial recipe pack — a bounded, inert, content-addressed JSON document that names an admitted scene, one direction document, gallery axes, preview render requests, declared effects, and a temporal audit — then plan or run the built-in cinematic-world workflow. Use when a user asks an agent to direct a 3D scene, iterate camera/look/sequence variants, or review compiled direction before explicit selection.
metadata:
  internal: true
---

# Author a cinematic recipe pack

Slopcamera's extension surface is declarative: coding agents extend behavior by
authoring bounded JSON documents, never by registering code. The operation
registry is closed, and the `cinematic-world` workflow composes only registered
operations. A recipe pack is the complete input contract for that loop.

## Documents an agent authors

1. **Scene** (`slopcamera.spatial-scene`, schemaVersion 1) — the admitted world.
   Usually already admitted through `slopcamera scene` commands; the pack binds
   it by digest, it is never embedded twice.
2. **Direction** (`slopcamera.spatial-direction`, schemaVersion 1) — semantic
   beats, actions, camera coverage, and look intents. `projectDigest` must equal
   the scene's canonical SHA-256. Character and subject references are semantic
   ids (`hero`), which compile against `entity_hero`-style scene entities.
3. **Recipe pack** (`slopcamera.spatial-recipe-pack`, schemaVersion 1) — the
   inert plan document:

```json
{
  "kind": "slopcamera.spatial-recipe-pack",
  "schemaVersion": 1,
  "packId": "recipe_<slug>",
  "sceneSha256": "<canonical scene sha-256>",
  "direction": { "<the direction document>" },
  "cameraId": "camera_main",
  "axes": ["camera", "lighting"],
  "previews": [{
    "name": "reel",
    "request": {
      "cameraId": "camera_main",
      "mode": { "kind": "beauty" },
      "selection": { "kind": "frame", "timeUs": 0 }
    }
  }],
  "effects": {
    "renderPlan": {
      "kind": "slopcamera.spatial-render-plan",
      "schemaVersion": 1,
      "quality": {
        "outputBytes": 8000000, "particleCount": 0, "pixelBudget": 2073600,
        "simulationSteps": 0, "texturePixelBudget": 2073600, "tier": "preview"
      }
    },
    "particleSystems": [],
    "simulationBakes": []
  },
  "temporalAudit": { "timesUs": [0, 2000000, 4000000] }
}
```

## Boundaries the pack enforces

- `axes` — 1 to 6 unique values from `performance | camera | lighting | materials | effects | sequence`.
- `previews` — at most 4 named render requests; each `request` is revalidated
  against the spatial render-request schema at workflow admission.
- `effects` — optional; when present it requires `renderPlan` (a declarative
  `slopcamera.spatial-render-plan` document with a quality tier and optional
  post-process stack) plus bounded `particleSystems`/`simulationBakes`. Each
  preview render first plans an integrity-bound effects document
  (`scene.effects.plan`) and renders with it.
- `temporalAudit` — optional; requires `cameraId`; `timesUs` bounds to 64 samples.
- `packId` — `recipe_` prefix, at most 64 slug characters.

## Invoking the workflow

The workflow input is `{ "pack": <pack>, "scene": <scene>, "source": { "path": "<repo-relative scene path>" } }`.
`pack.sceneSha256` must equal the canonical scene digest, so compute it first:

```sh
slopcamera scene inspect <scene.json> --json   # read sceneSha256
```

Plan is effect-free — it compiles the graph and reports requirements only:

```sh
slopcamera workflows plan cinematic-world --input input.json --json
```

Run executes the graph: inspection, direction check, compilation, one gallery
plan per axis, per-preview effects binding and render, and the temporal audit.

```sh
slopcamera workflows run cinematic-world --input input.json --json
```

## What the workflow never does

- It never selects or promotes a candidate. `spatial.project.select-candidate`
  and `iteration.select` are separate explicit operations a caller invokes
  after reviewing the plan outputs.
- Compiled direction stays `verified: false`; it is a proposal document, not
  applied state.
- Planning and graph compilation execute no operations, read no credentials,
  contact no providers, and mutate no project state.

Files in this skill

  • AGENTS.md601 B
  • SKILL.md4.2 KB
  • agents/openai.yaml259 B

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.