What does Halftone Collage Motion do?
Divide the creative process into two phases: first, decode the visual into an editable specification, then build the animation from approved stills or transparent layers. Default to a locally-compatible workflow, faithfully recording every generator and renderer.
Select Delivery Scope
First, determine which layer the user wants:
- Analyze only: deliver a breakdown, universal prompt, and JSON visual specification.
- Stills only: complete the analysis and generate, then curate stills.
- Collage motion: after still approval, create a silent motion negative.
- Full video: use this skill as the visual phase, then follow video production contracts, subtitles, voiceover, digital humans, and archiving rules.
If a reference image is missing but the user explicitly asks to "analyze this image," first request the image. If only a brief description is provided, draft the specification directly without repeating known information.
Select Execution Path
Default to the local path:
- Stills or layers: load the
imagegenSkill and use Codex's built-inimage_gen. Do not default to switching to the OpenAI API, CLI, or other image generation models. - Simple layer-by-layer assembly: use
scripts/render_assembly.py. - Complex layouts, multiple transitions, or component-level timelines: load the
hyperframesSkill, adhering to this specification and source declarations.
Only when the user explicitly requests Arcads, Nano Banana, Seedance, or Kling, and the corresponding tool is callable in the current session, read references/provider-boundaries.md and enter the external path. Before uploading local brand images, product images, or portraits, you must explain the upload location, cost, and purpose, and obtain user confirmation.
Do not infer the actual provider based on prompts, filenames, or expected models. Models not actually invoked may only be described as "target prompts" or "compatible styles," never as the output source.
Default Production Pipeline (Hard Default)
Execute strictly in the following order:
- Decode the reference image or brief description; generate a breakdown, universal prompt, and JSON visual specification.
- Use Codex's built-in
image_gento generate a complete still for each scene independently. Keep brand images, product images, or portrait references visible across all related calls. - Copy the generated results from Codex's default output directory to the project's
assets/stills/. Do not let the project persistently reference files under$CODEX_HOME/generated_images/. - Present the complete still to the user. Do not generate animations, FFmpeg assembly plans, or formal videos until the user explicitly confirms the stills.
- After still confirmation, use Codex's built-in
image_gento generate each individual paper-cut element. When a transparent PNG is needed, first generate a solid-color chroma-key background, then call theimagegenSkill's specifiedremove_chroma_key.pyfor local matting, and pass the alpha pre-check. - Use the approved stills to lock the final composition, build the
assembly-plan.jsonwith transparent layers, and then use local FFmpeg or HyperFrames to assemble from an empty background piece by piece. - Complete media probing, keyframe checks, and a source manifest. Default to recording
actual_still_provider: codex-image-gen,actual_motion_provider: local-ffmpegorhyperframes,external_uploads: false.
If the user directly provides all stills or transparent layers, change actual_still_provider to user-supplied. Only fill in Arcads, Nano Banana, Seedance, or Kling when an external call actually occurred.
Phase A: Decode Reference
On first decode, fully read references/visual-contract.md and references/schemas.md. Output in the following order:
### Breakdown: medium and technique, color field, elements and cut edges, composition and layers, single-scene idea, text, safe zone, mood.### Universal Prompt: write a self-contained natural language prompt that does not rely on "the image above."### JSON Spec: copy fromassets/templates/collage-spec.jsonand fill in, keeping valid JSON.
Write observations as reproducible parameters. For example, write "black and white dot halftone, approximately 65 lpi, white thin outline, 6px soft shadow at bottom right," not just "has a paper feel."
For multiple reference images, decode each separately, then add a paragraph summarizing the common style. When the user asks to "keep the look, only change the color," only modify the corresponding fields, but re-output the complete JSON.
Phase B: Plan Scenes
Keep only one core idea per scene. Lock down for each scene:
- Background color and canvas size.
- Main elements, supporting elements, layers, and final coordinates.
- Text content, typography, and editability.
- Entry sequence from empty background to full composition.
- Avoidance zones for real people, digital humans, subtitles, logos, and platform UI.
Build the assembly plan from assets/templates/assembly-plan.json. For the final formal video, prioritize the layered mode, providing one PNG with an alpha channel for each element. The bands mode only cuts the entire still into strips that slide in; it may only be used for quick previews or when the user explicitly accepts this simplified version.
Phase C: Generate and Approve Stills
Load the imagegen Skill and use Codex's built-in image_gen to first generate or composite the complete still, then present it to the user for approval. Keep brand or product references visible across all relevant generation calls.
The local formal mode works in the following order:
- Initiate one independent built-in
image_gencall per scene to generate the full composition still; do not combine different scenes into one large image with a single call. - Copy the selected stills to the project's
assets/stills/, recording the final prompt and the role of reference images. - Display the stills and obtain explicit user confirmation. Without confirmation, only allow prompt adjustments and regeneration.
- After confirmation, generate the individual collage elements. Follow the chroma-key plus local matting path according to
imagegento obtain transparent PNGs, retaining the white keyline and soft shadow. - Calibrate the size, coordinates, and layers of the elements using the approved stills. By default, preserve text as editable text layers in HyperFrames, HTML, or Remotion.
- Export the scene still and layer manifest, then proceed to the animation phase.
If the built-in image_gen is unavailable or fails, do not automatically switch to the CLI. First inform the user that the CLI/API path requires OPENAI_API_KEY, and switch only after the user explicitly agrees.
If only one complete still is available, you may use the bands mode for a low-cost motion preview, but do not describe rectangular strip slide-ins as a true "element-by-element assembly."
Phase D: Local Layer-by-Layer Assembly
First validate the specification and plan:
python3 .claude/skills/editorial-collage-motion/scripts/validate_collage.py \
--spec /absolute/path/collage-spec.json \
--plan /absolute/path/assembly-plan.json \
--project /absolute/path/project
Then render the silent negative:
python3 .claude/skills/editorial-collage-motion/scripts/render_assembly.py \
--project /absolute/path/project \
--plan /absolute/path/assembly-plan.json \
--output /absolute/path/project/renders/collage-assembly.mp4
The animation must start from an empty solid-color field, with elements entering from the edges of the frame according to the plan and settling into their final positions. Do not use camera shakes, full-image drifts, or continuous zooms to fake assembly. After settling, leave enough readable time.
Phase E: QC and Provenance Declaration
Fully read references/qc-and-provenance.md and complete at least:
- Media probe: dimensions, aspect ratio, frame rate, duration, codec, and decodability.
- Spot-check: first frame, before and after each scene settles, mid-scene, and last frame.
- Checks: starts from empty field, enters one by one, shadows stable, text readable, safe zone, and visual density.
- Ensure the
render-manifest.jsonis consistent with the actual output, input plan, and generators.
In the production notes, explicitly write:
actual_still_provideractual_motion_providerexternal_uploads- Whether it is a
compatible_renderand which visual specification it is compatible with
When the Arcads MCP has not been called in the current session, consistently use "Arcads style compatible render" or "editorial collage local render," and do not write it as a native Nano Banana or Seedance output.
Resources
references/visual-contract.md: style vocabulary, hard visual contracts, and counterexamples.references/schemas.md: visual specification and assembly plan field descriptions.references/provider-boundaries.md: boundaries and disclosure rules for local and external generation paths.references/qc-and-provenance.md: delivery QC and provenance recording.assets/templates/collage-spec.json: reproducible visual specification template.assets/templates/assembly-plan.json: reproducible local assembly plan template.assets/templates/assembly-plan-bands.json: strip preview template for use only when a complete still is available.scripts/validate_collage.py: pre-check of spec, plan, and asset materials.scripts/render_assembly.py: FFmpeg deterministic layer-by-layer or band assembler.