Skip to content
Back to search

story-shorts

SkillWorks with Claude CodeGitHub

Use when making faceless vertical (9:16) animated multi-part story videos, e.g. a zombie POV series, with Claude writing the story, Codex generating images, ElevenLabs narration and ffmpeg doing the montage. No paid video model is needed.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/onursamett--story-shorts-skill/badge.svg)](/listings/onursamett--story-shorts-skill)

HTML

<a href="/listings/onursamett--story-shorts-skill"><img src="/api/listings/onursamett--story-shorts-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

cp -r story-shorts ~/.claude/skills/story-shorts   # folder name = skill name

Pillow in a virtual environment (if your system Python is locked down):

python3 -m venv .venv && . .venv/bin/activate && pip install Pillow

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 65 lines
  1. ---
  2. name: story-shorts
  3. description: Use when making faceless vertical (9:16) animated multi-part story videos, e.g. a zombie POV series, with Claude writing the story, Codex generating images, ElevenLabs narration and ffmpeg doing the montage. No paid video model is needed.
  4. ---
  5. # story-shorts
  6. Pipeline: story bible -> outline per part -> character sheets -> anchored location images -> narration -> montage -> frame QA -> caption.
  7. Scripts live in `scripts/` (path below is relative to this skill folder), templates in `templates/`.
  8. Work in a project folder per part, e.g. `series/part1/` containing `scenes.json`.
  9. ## 1. Story bible first
  10. Copy `templates/story_bible_template.md` and fill it in before anything else:
  11. - a map of locations with the only legal routes between them,
  12. - ONE timeline with the lighting for each part (state day/night in every image prompt),
  13. - a cast with fixed one-line descriptions (hair, clothes, one unique accessory).
  14. Every later step must obey this file. Do not start outlining until the user approves the bible.
  15. ## 2. Part-by-part outline
  16. - Each part starts exactly where the previous one ended (first line = short "previously").
  17. - Explain every move between places; never jump location or time of day.
  18. - 1-2 locations per part, 60-90 s (12-16 scenes), ends on a cliffhanger.
  19. - Mostly medium/wide shots; at most 2 extreme close-ups per part.
  20. - New accounts: no A/B poll in the first parts. Add an A/B choice card only from later parts and pick the winner yourself until comments arrive.
  21. ## 3. Character reference sheets (6 expressions)
  22. Put one entry per character in `"refs"` of scenes.json: a full-body front view plus six head close-ups with clearly different expressions (e.g. smile, bored, confused, terrified, crying, determined). Generate them first and let the user approve them:
  23. `python3 scripts/gen_images.py <folder>` generates refs first (they are the first jobs).
  24. Engines: Codex (default, ~1 min/image) or `--engine gemini` (Antigravity CLI, a few min/image). If Codex reports a usage limit, switch the remaining files to Gemini, or split the list: one Codex run and one Gemini run on different files can run at the same time (never two Codex runs).
  25. Put the fixed descriptions in `"cast"`: names in prompts are expanded to them automatically.
  26. ## 4. Anchor scenes (consistent rooms)
  27. Give every image a location id: `"loc": "classroom"` (or one scene-level `"loc"`). For each location the FIRST generated frame becomes the anchor and is automatically passed as an extra reference to every later image of the same location, so layout, furniture and lighting stay the same. Use `"anchor": false` on an image that must not use/become an anchor (extreme close-ups, window views, a deliberately different angle of lighting). Make the first frame of each location a clean wide establishing shot, and check it before generating the rest. Check with `--dry-run` first: it prints each job with its resolved refs and anchors and calls nothing.
  28. ## 5. Shot list (2-3 images per narration line)
  29. For every image prompt state explicitly:
  30. - camera position (wide from the back of the room, medium from the front, over the shoulder...),
  31. - location and light ("bright late-morning daylight, NOT night"),
  32. - a specific facial emotion,
  33. - hands simple or out of frame (hands are where image models fail).
  34. Events seen from another place must be framed from the viewer's position, e.g. "seen from INSIDE the classroom through the window, window frame visible in the foreground". Do not draw an event as if the camera stood next to it when the hero is somewhere else.
  35. Scene fields: see `templates/scenes_example.json` (`images`, `narration` or `lines`, `chat`, `widgets`, `fx`, `text`, `accent`, `focus`).
  36. ## 6. Narration
  37. One narrator reads everything, including phone messages (the chat card appears on screen while the line is read; use `lines` with `"show": k` to reveal k messages). Avoid violent words such as "bites" or "blood": the TTS content filter rejects them (`voice.py` warns about a few). Rephrase ("hurt", "wound", "scratch").
  38. `export FAL_KEY=...` then `python3 scripts/voice.py run <folder> [--voice Brian]`. Try voices with `python3 scripts/voice.py test "sentence" Brian Daniel`.
  39. ## 7. Montage
  40. `python3 scripts/render.py <folder>` (Pillow needed; see README for a venv). Output: `<folder>/<foldername>.mp4`. Options: `--speed 1.12`, `--pill`, `--out`.
  41. Features: 2 s camera shots, 3-word Impact captions with yellow active word, big keyword at the top, chat card, widgets (counter, bar, pulse, tag, options), fx alarm/flicker, vignette, click-free PCM audio.
  42. ## 8. Frame QA before delivering
  43. Look at every image (contact sheet or one by one) and at 2-3 frames of the final video. Check: lighting matches the timeline, location matches the scene, no duplicate frames, no text/letter artifacts, characters match the sheets, hands, nothing sudden like day <-> night. Regenerate bad ones by name: `python3 scripts/gen_images.py <folder> k3_2.png`, then re-render. Never send a video you have not looked at.
  44. ## 9. Caption / post
  45. Fill `templates/post_caption_template.md` per part: hook question, 2-3 concrete details, question to the reader, cliffhanger, follow call, 5-8 hashtags, cover text.
  46. ## Known pitfalls
  47. - Parallel Codex runs mix up output files: always sequential; the script md5-dedupes and retries once.
  48. - `codex exec -i X` swallows the prompt: pass images as `--image=path` (the script does).
  49. - ElevenLabs timestamps come in segments that can split a word: `render.py` merges them.
  50. - The DIN-style font lacks the "·" glyph: the pill widget draws the dot itself.
  51. - Image models drift: repeating the location description and the light in every prompt is cheaper than regenerating.
  52. - Do not generate in the same folder from two jobs at once.
  53. - Generating images is slow (about 1 min each): run it in the background and tell the user.
README.md · README · 50 lines
  1. # story-shorts
  2. A Claude Code skill that turns an idea into faceless, vertical (1080x1920) animated multi-part story videos, for example a zombie POV series. Claude writes the story bible, outline and shot list; Codex generates the images; ElevenLabs (through fal.ai) narrates; ffmpeg + Pillow do the montage with camera moves, word-by-word captions, chat cards and info widgets. No paid video-generation model is used.
  3. ## Requirements
  4. - Claude Code
  5. - Codex CLI, logged in (image generation runs through `codex exec`), and/or
  6. - Gemini via the Antigravity CLI (`agy`, signed in) for `--engine gemini`: slower (a few minutes per image) but a
  7. good fallback when Codex hits its usage limit. Images are read from `~/.gemini/antigravity-cli/brain/`.
  8. - `ffmpeg` and `ffprobe` on PATH
  9. - Python 3 with Pillow (for `render.py` only)
  10. - A fal.ai key in the `FAL_KEY` environment variable (for `voice.py` only)
  11. ## Install as a Claude Code skill
  12. ```bash
  13. cp -r story-shorts ~/.claude/skills/story-shorts # folder name = skill name
  14. ```
  15. Pillow in a virtual environment (if your system Python is locked down):
  16. ```bash
  17. python3 -m venv .venv && . .venv/bin/activate && pip install Pillow
  18. ```
  19. ## Quick start
  20. ```bash
  21. mkdir -p series/part1 && cp templates/scenes_example.json series/part1/scenes.json
  22. python3 scripts/gen_images.py series/part1 --dry-run # job list with refs and anchors, calls nothing
  23. python3 scripts/gen_images.py series/part1 # sequential, resumable (only missing files)
  24. python3 scripts/gen_images.py series/part1 --engine gemini k8_1.png k8_2.png # Gemini for some files
  25. export FAL_KEY=... # never commit this
  26. python3 scripts/voice.py run series/part1 --voice Brian
  27. python3 scripts/render.py series/part1 # -> series/part1/part1.mp4
  28. ```
  29. Image styles: `--style anime2` (default), `anime`, `pixar3d` (or env `STORY_STYLE`). Image model: `--model` or env `CODEX_IMAGE_MODEL`. Narrator voice: `--voice` or env `ELEVEN_VOICE`. Render speed: `--speed` or env `STORY_SPEED`. Fonts: env `FONT_IMPACT`, `FONT_DIN`, `FONT_SANS`.
  30. ## scenes.json
  31. See `templates/scenes_example.json`. Top level: `cast`, `refs`, `chat_me`, optional `voices`, `scenes`. Scene: `narration` or `lines`, `images` (`file`, `prompt`, `refs`, `loc`, `anchor`, `setting`), `loc`, `setting`, `chat`, `widgets`, `fx`, `text`, `accent`, `focus`. Timing files keep their original names, `kelime_zamanlari.json` and `satir_zamanlari.json`.
  32. ## Costs
  33. - Voice: about $0.10 per 1000 characters on fal.ai ElevenLabs v3 (a 75 s part is roughly 900-1200 characters).
  34. - Images: generated through your ChatGPT plan via Codex; no per-image fee but it counts against plan limits.
  35. - ffmpeg/Pillow: free.
  36. ## Limitations
  37. - No true character animation: motion is camera pans/zooms and cuts between stills.
  38. - Image generation is sequential, about 1 minute per image; a 14-scene part with 2-3 images per scene takes roughly 30-45 minutes.
  39. - Consistency depends on reference sheets and anchors; always check frames before publishing.
  40. - Narration uses one voice for everything.
  41. ## License
  42. MIT, see [LICENSE](LICENSE).

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.