pvthiendeveloper/shared-claude-skills

Stickman video (người que)

Pipeline script: S=~/.claude/skills/stickman-video/scripts/stickman.py (Python stdlib + ffmpeg + PIL). Run python3 $S --help. References (read when needed):

• references/prompt-library.md: style presets, character locks, scene/motion prompt rules, script rules. Read before writing plan.json.
• references/research.md: what works, niches, tools, platform rules (from 50 tutorials).
• references/video-notes.md: raw per-video notes with verbatim prompts (grep by channel or keyword).

Projects live in ${STICKMAN_HOME:-~/stickman-videos}/<slug>/ (one folder per video): plan.json (source of truth) · audio/ · refs/ (character sheets) · images/ · clips/ (optional animated clips) · work/ · final/<slug>.mp4 + .srt · prompts.md.

Reply to the user in Vietnamese, short STE sentences. Run autonomously with sensible defaults; ask only for the topic if none is given.

Defaults (override when the user says so)

Item	Default
Format	16:9 long-form. "short", "tiktok", "reels" → 9:16
Length	long-form 3–5 min for a first test; Shorts 30–45 s
Language	Vietnamese
Voice	Gemini TTS Charon (works for vi and en). Better Vietnamese: ElevenLabs eleven_v3 with a Vietnamese voice: set voice.engine=elevenlabs and voice.voice_id in plan.json. Preview/fallback: {"engine":"say","voice_name":"Linh"} (offline macOS)
Style	doodle preset; comedy → rico-bw; motivation → paper; satire → mspaint; family/moral VN → pencil
Images	gemini-3.1-flash-image (Nano Banana 2). --model gemini-3-pro-image for higher quality
Music	none unless asked or a track is given (--music, -20 dB under voice). Use licensed / royalty-free tracks only and keep their credit lines.

Steps

1. Init. python3 $S init <proj> --aspect 16:9|9:16 --lang vi|en.

2. Write plan.json (you, Claude). Read prompt-library.md first.

   • title, slug, style (preset text), characters: [{id, lock}] (or {id, ref: "<image path>"} to use the user's own character image), voice, captions.
   • scenes[]: {id: "s01", narration, image_prompt, chars: [ids], motion, image_text?, motion_prompt?, pause?}.
   • Write the script voice-first: one sentence or beat per scene. Long-form 3–6 s per scene (≈10–20 words vi). Shorts ≤ 12–15 words.
   • Hook in the first 1–3 scenes. Open question every 20–30 s. Simple words.
   • image_prompt = one frozen moment, English, one paragraph. No text in images unless image_text (1–3 English words).
   • Add a no-character establishing shot every 4–6 scenes. Alternate motion.
   • ElevenLabs v3 accepts audio tags in narration ([curious], [whispers], [laughs]). They are removed from captions automatically.

3. Check. python3 $S check <proj>. Split every scene the script flags as long.

4. Voice. python3 $S tts <proj>. This writes per-scene wav files with edge silence trimmed, plus timing.json. Listen to 1–2 files (afplay) only if the user asks. Redo one scene with --only s07.

5. Character sheets. python3 $S sheet <proj>, then python3 $S contact <proj> --refs. Look at the sheet image. Re-roll with --force if the views do not match, have extra limbs, or the wrong style.

6. Scene images. python3 $S images <proj>, then python3 $S contact <proj>. Look at every contact sheet.

   • Re-roll a bad scene: edit its image_prompt if needed, then images <proj> --only s05,s12.
   • Bad means: wrong character design, garbled text, extra limbs, duplicate character, off-style (3D or photo), or the image does not match the narration. 6b. No image API? Draw the images yourself (free, offline). scripts/svgdraw.py is a code-drawn doodle library. It has these parts:
   • stickman(): 19 poses, expressions, outfits, hair, hats (crown, tiara, bicorne, mask…), dress.
   • pug(): sit, stand, lie, back.
   • About 40 props: furniture, sun, moon, ship, castle, pagoda, fridge, bowl…
   • Backgrounds: room, outdoor, dusk, night, palace, sea, plain.
   • PNG output via headless Chrome.

   Write <proj>/draw_scenes.py with one function per scene. Example: examples/cho-pug/draw_scenes.py. Run it, then check the result with contact.

   • Close-up of a sitting pug at scale s: pass y ≈ 560 + 123*s so the head stays in frame.
   • Add new characters or props to svgdraw.py when a topic needs them.

7. AI motion: use the cheapest mix by default. The first 5 scenes plus every 4th scene get an AI clip. All other scenes use the Gemini still with free Ken Burns and line boil.

   • Tested and not recommended: AI clips for every scene (a 2-minute video cost ~130 Kling credits). Code-drawn canvas animation looked too plain. Local Wan 2.2 on an M3 Pro was broken and slow.

   • Run python3 $S aiplan <proj> (--hook 5 --every 4). It lists the scenes that get an AI clip: the first 5, then every 4th scene. It also prints the credit estimate. For 35 scenes that is 12 clips: ~24 credits on Seedance 2.0 Mini 480p, or ~45 on Kling 3.0 at 3–4 s.

   • Before any paid generation, tell the user the clip count and the total credit cost. Wait for a clear yes.

   • All other scenes get free motion in render: Ken Burns plus hand-drawn "line boil" (3 warped copies cycled at 8 fps).

   • Render the mix: render <proj> --ai-hook 5 --ai-every 4. Without --ai-hook, every clip present is used.

   • Model prices (Higgsfield, 16:9, silent):

     Model	Setting	Credits
     Seedance 2.0 Mini	480p, 4 s	2
     Seedance 1.5	480p	2.4
     Grok Imagine 1.5 Lite	480p, 3 s	3
     Kling 3.0 std	3 s	3.75
     Kling 3.0 std	4 s	5
     Kling 3.0 std	6 s	7.5

     Kling 3.0 is proven on this style. Seedance Mini is untested; test 1 clip first.

   • Local models on this Mac (M3 Pro 36 GB) were tested and rejected. Wan 2.2 TI2V-5B GGUF took 16 min for a 2 s clip at 832×480, and moving parts broke into blocks. Do not reinstall.

   • Use the Gemini API (reliable, ~35–45 s per clip): VEO_DUR=4 python3 scripts/veo.py images/<id>.png "<motion_prompt + style-lock tail>" clips/<id>.mp4.

   • VEO_DUR is 4 or 6 (scene ≤ 4 s → 4). Model: veo-3.1-lite-generate-preview.

   • Run clips one at a time. Parallel requests hit HTTP 429 quota. On a 429, wait 60 s and retry.

   • Google Flow in Chrome also works but is slow, and its agent asks for 10 credits per clip. Click Approve, never "Always approve".

   • Manual alternative: The user may want motion in the first 5–10 scenes.

   • Run python3 $S export <proj> to write prompts.md.

   • The user animates those scenes in Google Flow / Veo / Grok (start frame = images/<id>.png).

   • The user saves the clips as clips/<id>.mp4. render uses them automatically. Video audio is dropped.

8. Render. python3 $S render <proj> --ai-hook 5 --ai-every 4 [--music <file>] [--music-db -20] [--no-captions].

   • Then extract 4 frames and look at them: ffmpeg -ss <t> -i final/<slug>.mp4 -frames:v 1 work/f.png.

9. Deliver. Give the user these items:

   • The path to final/.
   • The duration.
   • What you chose: style, voice, scene count.
   • A YouTube title, a description and 3–5 hashtags.
   • A reminder to set "Altered or synthetic content = Yes".
   • Offer to deliver the file through any messaging or file-sharing tool the user has.

Revisions

• Change one scene's text: edit narration, then run tts --only <id> and render.
• Change one picture: edit image_prompt, then run images --only <id> and render.
• Change the voice: edit voice, then run tts --force and render.
• Change the style: edit style, then run sheet --force, images --force and render. This re-pays for every image, so tell the user first.

Rules

• Original scripts only. Never rip frames or scripts from other channels (reused-content demonetisation).
• Keys come only from env (GEMINI_API_KEY, ELEVENLABS_API_KEY). Never print them.
• Failures:
  • HTTP 402 or 429 from Gemini, or 401 from ElevenLabs: stop. Tell the user that the key or credits need attention. You may offer the say voice as a preview.
  • Never loop on retries.
• Cost: every image is one paid API call. A 10 min video is about 150 images. Before a long-form run over 60 scenes, tell the user the scene count.
• Captions are timed by character share inside each scene, not by forced alignment. Timing is accurate per scene but approximate inside the scene.