youtube-video-creator-skill

YouTube video creator — narration → Manim → DrawCMS

Turn a folder of narration scripts (one markdown + one mp3 per chapter) into a chapter-per-video series plus one combined video, perfectly synchronised to the narration:

my-series/
├── narration/            # one .md + one .mp3 per chapter
├── .drawcms/             # DrawCMS diagram specs + build receipts
├── video/                # the pipeline (from templates/video/)
│   ├── segments.py       # content spec: which beats cover which paragraphs
│   ├── timing.py         # paragraph → seconds allocation from real audio
│   ├── beats.py          # SegmentScene: renderers, animations, diagrams
│   ├── scenes.py         # one Manim scene class per segment
│   ├── theme.py          # palette + layout bands (rebrand this)
│   └── build.py          # diagram | schedule | render | mux | concat
├── schedule.json         # validated timeline (generated)
├── manim_build/          # rendered scenes (generated)
└── final/                # 01-*.mp4 … + full.mp4 (generated)

When to use

• The user has (or wants to write) narration markdown + mp3 files and wants videos built from them.
• Rebuilding an existing series after narration fixes, content edits, or style changes.
• Adding chapters, re-timing beats, or adding DrawCMS diagrams to a series.

Do not use this for: live-action editing, screen recordings, or audio-less videos — this pipeline is specifically script+narration → motion graphics.

Prerequisites

Tool	Needed for	Notes
Python ≥3.10 + venv/uv	everything	uv venv .venv && uv pip install --python .venv/bin/python manim
Manim Community (0.22.x)	rendering	manim binary inside .venv/bin/
ffmpeg + ffprobe	durations, mux, concat	on PATH
Node + DrawCMS CLI	diagrams (optional)	~/.agents/skills/drawcms/bin/drawcms.mjs; override with DRAWCMS_CLI
faster-whisper	narration verification (optional but recommended)	uv pip install faster-whisper

Fast path

Run everything from the project root (the folder holding video/).

0. Scaffold. Copy templates/video/ → video/, templates/fixtures/ contents → narration/, templates/.drawcms/ → .drawcms/. Create the venv and install manim. Prove it works before touching content: python templates/smoke_test.py (from this skill's folder).
1. Verify narration. Transcribe every mp3 with faster-whisper and diff it against the markdown word-for-word; fix mismatches before building visuals. → references/narration-sync.md
2. Author the content spec. For each segment, split the script into paragraphs and give every beat a paras: [start, end) range. The ranges must tile the script exactly — build.py schedule enforces it. Beat text must be shorter than the narration it covers. → references/manim-beats.md
3. Validate timing. .venv/bin/python video/build.py schedule → all segments OK, total duration = sum of mp3s. Fix any tiling error now.
4. DrawCMS diagrams (optional). Author typed JSON specs in .drawcms/, build receipts, drive them with architecture beats that highlight subsets per beat. → references/drawcms-diagrams.md
5. Render one scene first (build.py render intro --force), inspect frames, then render the rest. → references/verification.md
6. Verify. Beats probe, frame grid, full-res crops of every diagram beat, bounds checks. → references/verification.md
7. Mux + concat. build.py mux --force then build.py concat (concat always rebuilds). Check final durations with ffprobe.
8. Iterate. After any code/content change: render <ids> --force → mux --force → concat. Renders are skipped when outputs are current, so a restart is always safe — but mux must be forced after a re-render or you ship stale video with new audio.

Detailed command reference and layout rationale → references/workflow.md.

Invariants (never violate these)

1. Paragraph ranges tile the script. [0,a), [a,b), … [z,n) with no gaps or overlaps. Timing derives from this; a gap = a silent stretch with no visuals, an overlap = double-counted seconds.
2. Beat text ≤ narration. The narrator says more than the card shows. Cards that read longer than the audio feel wrong even when sync is exact.
3. No progress bar. The narration audio is the timeline; a bar adds noise and dates the video.
4. Centered, sized text. Multi-line text is centered per line (Manim left-aligns lines by default — _text/_center_lines handle this). Keep the type scale in references/text-animation.md.
5. Animations stagger; they never pop. Entrances are progressive (LaggedStart / scaled FadeIn), fit inside ~34% of the beat, and the rest of the beat is a hold. The final beat fades out with a ~0.6s tail.
6. Diagrams stay in their band. Heading + diagram fit below the header rule (_r_architecture scales them in). Node labels must fit inside their boxes (auto-fit in _node_label); no mid-word breaks.
7. Edge labels only when ≤4 edges are highlighted (built into _arch_style); more than that and the diagram becomes unreadable.
8. DrawCMS receipts are the source of truth for diagram topology and layout — never hand-edit .built.json; edit the spec and rebuild.
9. Force mux after re-render. render --force alone leaves final/ stale.

Skill layout

Path	Purpose
references/workflow.md	End-to-end pipeline, commands, iteration loops
references/narration-sync.md	ASR verification, paragraph rules, timing math
references/manim-beats.md	Beat vocabulary, scene architecture, adding beat types
references/text-animation.md	Type scale, centering, entrance-animation rules
references/drawcms-diagrams.md	Spec authoring, integration, layout rules
references/verification.md	Probes, frame grids, bbox checks, failure table
templates/	Drop-in pipeline code + fixture narration + smoke test
agents/	Adapters for OpenCode, Claude Code, Cursor, AGENTS.md
install.sh	Symlink this skill into ~/.agents/skills and ~/.claude/skills