Outlined story cartoons

---
name: faceless-story-cartoon
description: Produce faceless YouTube survival or serialized drama videos with consistent outlined cartoon puppets, painted locations, word captions, day counters, fast cinematic cuts, narration, sound, and verified final video.
---

## Public package setup

Export `KIT` to your extracted kit directory and `SKILL` to this extracted style directory.
Read README.md for dependencies, missing assets and setup. Example media and private review
material mentioned below are not bundled; create your own input/output paths.

Read [SPEC.md](SPEC.md) before writing the script or storyboard. It records the local reference study, palette, fonts, dimensions, cinematography and transition rules.

Use this folder as the render root. KIT is `$KIT`; PY is `$KIT/.venv/bin/python`. The kit reads credentials itself. Never read its env file, modify the kit, or use yt-dlp. All generated output stays inside this skill. Network calls go through kit scripts. Never borrow reference video, branding, lines or characters for output.

## Production

Write an original script: `Zombie Outbreak: 30 Days [specific confinement]` is the survival title formula. First-person narration, immediate concrete danger/safety assumption, first day action, next day reversal, unresolved reveal. A 15-second opening is 37–40 words at about 150wpm; longer serialized episodes can use the measured 179wpm if comprehensible. Do not claim the 15s hook contains thirty days. Day counters follow narration. Dialogue and narration use the same Scribe-aligned caption track; optional speaker labels belong in the safe upper region.

One command produces `final.mp4`, `thumbnail.png` and an actual QC result:

```sh
./build.sh path/to/script.txt --topic 'Zombie Outbreak: 30 Days Trapped in an Aquarium' --out examples/output --quality max --target-duration 15 --storyboard path/to/storyboard.json
```

Omit `--storyboard` for automatic script-specific painted backgrounds and sentence/reaction planning. This works as a starting point, but inspect semantic accuracy and replace beats with an authored storyboard for production. Automatic art never substitutes aquarium backgrounds for an unrelated story. Use `--stage voice|plan|stills|render|audio|qc` to resume; existing voice, words and music are cached. Delete only the specific output you intend to regenerate. Do not reuse timings from a changed voice.

The driver runs MiniMax voice → optional target-duration pacing → ElevenLabs Scribe → planner → stills → standard/max preview and vision QC → deterministic SS render → thumbnail → continuous music → mix → final kit QC. Failed QC stops the command. No pass is implied by successful rendering.

For a new art direction, author a JSON storyboard with title, location, character settings and `beats`. Each beat specifies `at` (an exact spoken phrase; first beat starts at zero), background (render-root-relative image), environment, size, angle, camera, day, hero, optional zombie/prop/title/sound. Example schema and working beats: [examples/test/storyboard.json](examples/test/storyboard.json). Generate background plates via `$PY $KIT/py/gen.py image 'prompt' output.png --model seedream --aspect 16:9`, then view them and log prompt/model/path. Assets must depict coherent attached machinery and true openings. The renderer accepts arbitrary plate paths and timelines; aquarium props are optional modules, not required scenery. For other people, put named profiles in storyboard `cast` and per-shot placements in `actors: [{"id":"friend","character":"friend","x":700,"y":900,"scale":1,"pose":"point"}]`; every actor uses the same parametric generator and is tracked by QC.

```sh
$PY scripts/plan.py script.txt words.json --storyboard storyboard.json --out examples/output/timeline.json --quality max
node $KIT/js/render.mjs --root . --page scene.html --data examples/output/timeline.json --ss 2 --still 2.0 --out examples/output/check.png
node $KIT/js/render.mjs --root . --page gallery.html --ss 2 --still 0 --out characters.png
```

## Art and edit constraints

Use the parametric rig in `modules/character.js`: skin, shirt, pants, hat and hair colors; headwear beanie/cap/none; build slim/ripped; expression worried/fear/surprised/happy/angry; poses idle/point/reach/guard/fear/hold; facing left/right; satchel accessory or none. `characters.png` proves six configurations including a ripped red-tank character. Keep the lead's settings fixed across shots. Set full-body feet to a visible floor baseline; close-ups can intentionally crop the torso, never eyes or headwear.

Plan 4–5 cuts per 10s, mostly 1.5–2.7s shots. Alternate wide geography, medium action, close face, extreme-close hand/mechanism. At least one wide and two motivated inserts per 15s. Set 2–5% deterministic camera drift per shot, modest pans; change angle and framing at cuts. Hard-cut on clauses or significant words. A 0.22s desaturated dissolve may mark a day jump. Titles fade over 0.2s in and 0.28s out. Never toggle a title away mid-shot. Keep the actor clear of solid doors and props; don't slide the head into a gate.

Captions stay in the bottom reserved band; day/location share a non-overlapping top rail. Typography uses system Comic Sans MS Bold, Chalkboard SE and Arial Black; no font files are redistributed. World-camera transforms never affect caption placement. `data-qc` covers text; `__qc` tracks actors across cuts. For new labels use measured boxes and don't anchor them across faces. See SPEC for all dimensions.

Sound: continuous quiet ominous ambience, no vocals, minimal drums. Latch click at the latch; electrical downlifter at power loss; knock at glass contact. Cues use the kit's bank and peak timestamps, generally 10–14dB below voice. Mix with the kit at -14 LUFS. Avoid automatic whooshes on every cut.

Draft uses SS1 and six particles; standard uses SS2 and 15; max uses SS2 and 30 with full glass treatment. All levels use the same coherent art, timing, layout and correctness. Preview standard/max before final rendering. Record measured render cost per video second in the output notes; don't extrapolate unmeasured quality levels as facts.

## QC loop and delivery

View stills, including a six-moment sheet beside matching reference shot sizes. Fix the largest visual differences in reusable modules or planning. Run the kit QC on the mixed final with `--qc silent.qc.jsonl --timeline timeline.json --words words.json --style 'Outlined 2D story cartoon'`. Read report and flagged frames full-size. A disputed finding remains open. Fix and rerender until exit 0; `ignore.json` is only for an intended effect that has been personally inspected and documented, never for evading a real defect. Max requires a second vision pass on flagged spans after fixes.


For a local code fix after a full render, replace a frame-aligned span without retiming the episode:

```sh
$PY scripts/repair.py examples/output --from 5.8 --to 8.8 --quality max
./build.sh examples/output/script.txt --out examples/output --quality max --stage audio
./build.sh examples/output/script.txt --out examples/output --quality max --stage qc
```

Include any following dissolve that contains the changed previous shot. Repair validates total frame count and replaces the span's QC telemetry. It retains a before copy for audit. Re-export reference comparisons and final inspection frames after repair. For quality timing, `scripts/benchmark.py examples/output` measures two-second draft/standard samples and includes full-render max metrics.