sizzle

name: sizzle description: Scenes folder + style → evaluated, validated, rendered sizzle reel video.

/sizzle - Sizzle Reel Pipeline

Full AI cinematography pipeline: load scenes → analyze → plan → evaluate → validate → render video.

Command Interface

/sizzle <scenes-dir> --style <name> [options]
  --style <name>        Style pack (required). One of:
                        prestige, energy, dramatic, minimal,
                        intimate, corporate, kinetic, fade
  --output <path>       Output path (default: renders/sizzle-{style}-{timestamp}.mp4)
  --dry-run             Generate manifest JSON only, skip render
  --skip-evaluate       Skip quality evaluation step
  --skip-validate       Skip guardrails validation step
  --sequence-id <id>    Custom sequence ID
  --verbose             Print per-scene analysis details

Parameters

Execution Flow

When /sizzle is invoked:

1. Read Scene Files

Load all *.json scene files from the specified directory. Validate each against the scene format spec.

2. Analyze Scenes

Call analyze_scene MCP tool on each scene to classify:

• Content type (typography, portrait, device-mockup, etc.)
• Visual weight (light, dark, mixed)
• Motion energy (static, subtle, moderate, high)
• Intent tags (opening, hero, closing, etc.)

3. Plan Sequence

Call plan_sequence MCP tool with analyzed scenes + style pack:

• Orders scenes by intent progression
• Assigns transitions per style pack rules
• Sets camera overrides per content type
• Assigns hold durations by motion energy

4. Evaluate Quality

Call evaluate_sequence MCP tool to score the manifest:

• Pacing — duration match to style pack hold durations
• Variety — shot size, content type, weight, energy diversity
• Flow — energy arc, intent progression, transition coherence
• Adherence — camera, transition, shot grammar compliance

Present the overall score (0-100) and per-dimension breakdown.

If the score is below 60, recommend reviewing the findings before proceeding.

5. Validate Guardrails

Call validate_choreography MCP tool to check safety bounds:

• Speed limits on camera moves
• Personality boundary enforcement
• Lens bounds (scale, rotation)
• Acceleration and jerk limits

If verdict is BLOCK: Must fix violations before rendering. Show the blocking issues and suggest corrections.

If verdict is WARN: Show warnings but allow proceeding.

If verdict is PASS: Continue silently.

6. Human Review

Before rendering, always present the manifest for review:

• Show scene order with durations
• Show transition types between scenes
• Show camera overrides
• Show the evaluate score and any warnings

Ask the user to:

• Approve — proceed to render
• Edit — modify the manifest JSON directly, then re-evaluate
• Re-plan — try a different style pack or adjust parameters
• Cancel — abort the pipeline

7. Render

On approval, run the CLI to produce the video:

node scripts/sizzle.mjs <scenes-dir> --style <name> [--dry-run] [--verbose]

Style Packs

Examples

Basic sizzle reel

/sizzle scenes/ --style prestige

Dry run for review

/sizzle scenes/ --style dramatic --dry-run --verbose

Fast pipeline (skip quality checks)

/sizzle scenes/ --style energy --skip-evaluate --skip-validate

MCP Tools Used

Rules

DO

• Always show the manifest before rendering
• Always run evaluate + validate unless explicitly skipped
• Present findings clearly with severity levels
• Allow the user to edit and re-evaluate
• Use --dry-run for first iterations

DO NOT

• Render without human approval
• Ignore BLOCK verdicts from guardrails
• Skip evaluation when the score matters (marketing, external use)
• Mix style packs within a single sizzle reel

Related Commands