agentleFS
shaharshashaharsha/claude-skills/skills/producing-animated-episodes/SKILL.md

producing-animated-episodes

Skill6 starsChanged 22 days agoMITView on GitHub

Use when making a narrated or voiced animated video with recurring characters - an anime, cartoon or comic episode, a short for a group chat, a story video about friends, family, a class or a team - from photos, a chat export, a story idea or a script, in any visual style, Hebrew/RTL or English. Also use when such a video has subtitles or cards too fast to read, music stopping early, the ending cut off, lines buried under music, mouths out of sync, a gesture before its line, or one face bigger than the others in a group shot.

  • Reads credentials
  • Installs packages

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 1 is distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Every gate is a stop: send the artifact, ask, wait.

    line 29
  2. Distinctive:

    Before each paid phase, give the cost estimate.

    line 85

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Producing animated episodes

Overview

A vertical, voiced, subtitled, animated episode: character art, panels, cloned or library voices, music, image-to-video clips and lip-sync, cut by an engine that times every shot from the real dialogue audio. One episode.json holds the whole project, and engine/episode.py runs every phase from it. The sibling skills do the provider work.

Use the engine; don't write your own edit scripts. Timeline, overlays, subtitles, right-to-left text, the mix, durations and lip-sync joins are already in engine/. It was proven by rebuilding a real 3.5-minute episode: identical frames, identical audio, and the exact length the original had missed. A hand-written edit rediscovers every bug in references/troubleshooting.md.

The audio is the master. Shot lengths, subtitle times and lip-sync parts come from the finished dialogue files. Never estimate a timing from text.

You can't watch or hear. You check frames, durations and numbers. The human judges likeness, voices, pronunciation, timing and jokes. Plan their gates and keep them.

Requires:

  • These skills next to this one: image-generation, elevenlabs-tts, generating-video-clips.
  • Python 3 with pip install -r <this-skill-dir>/requirements.txt (Pillow built with RAQM, numpy; OpenCV optional).
  • ffmpeg 7 or newer.
  • Fonts: styles and subtitles name font keys (bold, hblack, black in the example) that fonts maps to installed .ttf files. For Hebrew and Latin, Rubik and Heebo (free, Google Fonts) work well; check blocks on a missing one.
  • Keys in env vars, set in every shell that runs the engine: OPENAI_IMAGE_API_KEY, ELEVENLABS_API_KEY, GEMINI_API_KEY, FAL_KEY.

To see the whole pipeline work with no API calls, run python3 <this-skill-dir>/engine/episode.py /tmp demo, then check and build the demo project it creates.

The pipeline and its gates

Every gate is a stop: send the artifact, ask, wait. Commands are python3 <this-skill-dir>/engine/episode.py PROJECT <command>. Costs measured 2026-10.

#PhaseCommandGate (the human)Cost
1Story and scriptscript.md from templates/script.template.mdthe script, and every sensitive topic-
2Character bibletemplates/characters.template.md, then characters in episode.json--
3Faces, cards, promotechars face, chars card key=face.png, chars final keyslikeness, on one sheet of all cards~$0.30 per character
4Panelspanels draft, then panels final idsevery panel, on a sheet~$0.30 per panel
5Voicesone sample per voice (elevenlabs-tts)every voice, by ear, before the batchcents
6Lineslinesthe set by ear, with timestamps; names A/Bcents to a few $
7Music and SFXsoundthe musiccents
8Animaticbuild (no clips yet: stills with camera moves)pace and order, before any video spendfree
9Animationanimate, sheets, retakes with animate S --takes 2 + animate pick S Kmotion, from sheets~$0.10 per second of clip, per take
10Lip-synclipsync prep, face points in config, lipsync runspot-check frames~$0.13 per second
11Edit and deliverbuild, sheet, deliverthe human watches the mp4free

Commands, inputs, outputs and failure handling for each phase: references/pipeline.md.

Writing the script

Rules that held up (details and a worked example: references/script-writing.md):

  • A hook in the first 3-4 seconds (a cold open on the problem), then the title. Under 30 seconds, put the title and end card as overlays on story shots and skip the montage.
  • Every recurring character speaks at least once. An intro montage, one word each in their own voice at about 1.2-1.5 s per card, is cheap and loved.
  • On-screen text must be readable: about 0.8 s plus 0.06 s per character; 1.2-1.5 s for a short card; 4-6 s for a final caption. Give cards a mindur.
  • Plant clues early and pay them off. Callbacks to a group's real running jokes land hardest.
  • Sensitive topics are the user's call. List every private detail you'd like to use (health, money, relationships, family, fights), ask about each, and use only what they confirm.
  • Lines of 12 words or fewer, numbers in words. On-screen text stays in normal spelling; pronunciation fixes go in pronounce.
  • Music and sound effects, even for narration. At least one music bed and a few effects (whooshes on cuts, a hit on the reveal); a silent bed sounds unfinished.

episode.json in one screen

templates/episode.example.json is a complete small episode; every key is in references/config-schema.md. The parts you edit most:

  • characters: display name, color, voice, and the prompt fields from the character bible.
  • speakers: voices that are never drawn, such as a narrator ({"display": null, "voice": "<id>"} gives subtitles without a name).
  • lines: id, speaker, text with audio tags; trim and level on one-word lines.
  • shots: image, lines ("02-3@-0.4" overlaps the previous line), pads, mindur, camera, fx, sfx, overlays, clip offset and speed.
  • music_cues: [cue, from shot, to shot, volume, offset]. The build warns when a cue is shorter than its window.
  • lipsync: which shots, which lines to skip, one face point per part.
  • prompts: override char_style and panel_style for a different look (3D cartoon, watercolor, photoreal).

Building and reviewing

  • Batch the fixes. A full build takes minutes. Collect every note from a review round (a line take, a pad, a clip offset), apply them all, then rebuild once.
  • Segments are cached and tracked. build re-renders exactly the shots whose inputs changed: a new clip or lip-synced version, a promoted panel, an edited line, shot or style. build --only S05,S06 forces those shots; --force re-renders everything.
  • A lip-synced shot is redone after any change to its clip, offset, speed or lines: lipsync prep S05 && lipsync run S05. Until then the build warns it is stale and uses the raw clip.
  • Regenerating an existing asset needs --force (sound theme --force, animate S05 --force); without it, existing files are kept and only gaps are filled.
  • Review with sheets. sheet puts 3 frames per shot on a page. When the human names a moment by time, extract and read those frames before answering (ffmpeg -ss T -i build/<output>.mp4 -frames:v 1 f.jpg).
  • check before every build: missing assets, keys, sibling skills, and overlay text wider than the frame.
  • Durations are verified at the end of every build. Audio shorter than video fails the build instead of cutting the ending.
  • Before sending: references/qa-checklist.md.

Working with the human

  • Send every artifact with your file-sending tool and offer to reveal it in Finder; don't just print a path. The final step is deliver and sending that file.
  • For audio, say what to listen for and where ("at 0:41, does Dani's name sound right?"). Never claim you heard it.
  • Pronunciation and voice choices are theirs: A/B files with timestamps (elevenlabs-tts).
  • Before each paid phase, give the cost estimate; after it, the running total.

When something looks wrong

SymptomFix
Cards or captions too fast to readraise mindur / post: ~1.2-1.5 s per short card, 4-6 s for the final caption
A title clipped at both edgescheck warns; lower the style's size or add wrap
Music stops before the section endsthe build warned: raise the cue's seconds, sound <cue> --force; or split the section into two cues
The ending is cut offcan't happen with the engine (exact-length mix, duration check); in hand-made muxes, see deck-to-video
A line is buried under musiclevel_clips.py --report on that line, then a firmer take; lower that cue's volume
A character points or acts before their lineclip_off past it, or split the shot and crop the first part with clip_cam
Mouths don't move with the wordslip-sync phase; if the wrong mouth moves, fix its face point
One face bigger in the group shotinstall OpenCV (faces are equalized), or mirror a card whose prop covers a neighbor
Hebrew renders reversed or unjoinedPillow without RAQM: check says how to fix it
A name or loanword mispronounceda pronounce entry, A/B by ear

More, with causes: references/troubleshooting.md.

Related skills

  • image-generation: character cards and panels ("Recurring characters and story panels").
  • elevenlabs-tts: voices, one-word lines, pronunciation, music and SFX.
  • generating-video-clips: clips, take sheets, lip-sync.
  • deck-to-video: the same ffmpeg lessons for slide videos.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.