morph-explainer-video

A Claude Code skill that turns a weird body fact, a "what if" question, or a finished script into a vertical (9:16) morph-explainer short — the deadpan anatomical / body-horror format where the subject deforms continuously against a flat cyan backdrop while a flat narrator reads ~165 words per minute over a wall-to-wall music bed.

Think "Zack D. Films style": cursed science shorts, "what happens if you did X to a human body" videos, built end-to-end with AI generation.

What it does

Given a fact, question, or script, the skill runs a full production pipeline on Unsora:

elicit (format, subject, length) → write script to template → [GATE 1: script approval]
  → create_voiceover (one unbroken take, ElevenLabs v3) → ffprobe the real duration
  → derive keyframe times & segment durations → [GATE 2: shot plan + credit estimate]
  → create_image K0..Kn (keyframe chain — each Ki is the reference for Ki+1)
  → create_video per segment (Seedance 2.0, image + lastImage pinned) → ffmpeg concat
  → create_music (mureka-7.5) → mix: bed + VO + 2 stings → −14 dB RMS master → final.mp4

It can also stop early and deliver just the script, or just one segment's prompt pair.

Key design decisions

• Voice first, picture second. The narration is generated as one take before any picture exists; every downstream duration is derived from the VO's real (ffprobe-measured) length, not the plan.
• One locked house style. The look never varies — only the script architecture does. Two formats ship: hypothetical (~117 words / ~43s, ends on a deadpan punchline) and true-story (~94 words / ~33s, ends on "now they plan to…").
• Cuts are the default grammar; morphs live inside a shot. Segments are pinned at both ends (image + lastImage), so the model purely interpolates. Between segments you cut — match cuts share a keyframe file, scene cuts get a fresh one.
• Nothing generates before the gates. Script approval is free (Gate 1); the shot plan plus a credit estimate must get an explicit yes before the first paid call (Gate 2).
• Mandatory calibration. One test image is generated and compared against a real reference frame before the full keyframe set — the style vocabulary is easy to miss and the failure is invisible from the prompt alone.
• Measured, not vibed. Shot lengths, words-per-minute, motion density, and the −14 dB master target all come from measurements of reference videos, and the doc records where earlier assumptions were wrong.

Installation

Copy the folder into your Claude Code skills directory:

# Project-level
cp -r morph-explainer-video .claude/skills/

# Or user-level
cp -r morph-explainer-video ~/.claude/skills/

Then in Claude Code, ask for a video in this style ("make a body horror explainer about…", "what if you swallowed a magnet — make it a short") and the skill triggers automatically, or invoke it directly.

Requirements

• Unsora MCP connector (create_image, create_video, create_voiceover, create_music, wait_for_*) — generation spends Unsora credits.
• ffmpeg / ffprobe on PATH for concat, mixing, and mastering.

Repository contents

morph-explainer-video/
└── SKILL.md   — the complete skill: pipeline, laws, phase-by-phase
                 instructions, failure table, and checklists

Note: SKILL.md references a references/ directory (house-style.md, pipeline.md, and the format templates) that is not included in this distribution. The skill's core logic is self-contained in SKILL.md; the reference files hold the locked style slots and exact API parameters.

License

No license specified — all rights reserved by the author unless stated otherwise.