warrior-clash-studio · Open Skill 공개 스킬
warrior-clash-studio
With this skill, you can enter [Character 1] VS [Character 2] or upload images to generate a video similar to a fighting‑game cutscene. This is a short‑video skill built for creating arcade‑style fighting game videos. Upload your custom character avatar, equip them with cool combat moves and battle animations to restore the fiery vibe of classic arcade fights. You can freely adjust battle tempo, hit effects and camera performances. No tedious editing needed. Easily recreate retro‑arcade combat atmosphere and produce high‑intensity fighting clips with stunning moves to make your very own arcade‑style game videos.
Manhua
28 13 40
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3656543/2026-08-25/4b5hhaww.gif
# warrior-clash-studio
## Reference Files
- **`references/fighting-rules.md`**: Default fighting-rules plan. Before Step 3 grids and Step 4 videos, read and follow this file (five-stage flow and HP) when the user has not requested other fight rules.
- `references/action-fx.md`: Action VFX.
- `references/ui-guide.md`: UI design.
- **`references/character-facing-rules.md`**: Left character (Character 1) faces right; right character (Character 2) faces left. Applies to Step 2, storyboard images 1–5, Videos 1–4, and Video 5 while the winner is still on their original side. Sole exception: storyboard image 6 / Video 5 last frame — winner may move to the **center of that cell** for Victory (not the center of the whole Grid B sheet). Read and follow first.
- **`references/language-rule.md`**: Reply-language rules. Read and follow first.
## Language Convention
- Reply in the language of the user's own input only. Do not judge language from SKILL.md, references, English templates, tool names, or button text. If the user's input mixes languages and the language is unclear, ask with **only** language buttons (e.g. `{"text": "中文"}`, `{"text": "English"}`), then stop and wait. Do not mix language buttons with style or winner buttons. Reading `references/language-rule.md` and that language-only `set_guidance` are the only actions allowed before the Step 0 confirmation copy.
- Startup order after the user sends input: (1) bind language from the user's own input; (2) Step 0 confirmation copy in that language; (3) style buttons only, wait; (4) Step 0.5 winner buttons only, wait; (5) only then may `search_web`, `generate_image`, `generate_video`, and other generation tools be called.
- Confirmation copy, button `text`, and stage notifications in this Skill are English templates: translate them into the bound language. Style-button meanings stay Retro Pixel / Modern 3D / Future Mecha; only labels are translated. Fighting-game UI in the picture (`READY? GO!`, `K.O.`, `Victory`) stays English.
## When to Use
- The user uploads two characters (images/descriptions) and wants a fighting-game versus video.
- UI lives **inside each 16:9 fight cell**: health bars upper-left and upper-right of that cell, plus portraits in that same cell. Character 1 left, Character 2 right.
## Steps
- **Image-generation channel lock (every `generate_image`)**: Non-empty `input_urls` (uploads, Step 2 versus, split tile, illogical cell, or any other image) → `model='i2i.hd.lite'` only. **Exception — Step 3 grid generation**: the two 4-panel storyboard grids (Grid A / Grid B) and their cell logic-redraws / regenerations use `model='i2i.hd.gpt'` (with the Step 2 versus as `input_urls[0]`) instead of `i2i.hd.lite`. No reference image (text-only, no `input_urls`) → `model='t2i.hd.lite'` only. Do not mix: no `t2i.hd.lite` with `input_urls`; no `i2i.hd.lite` with no reference. Forbidden: `t2i`, `t2i.hd`, `i2i`, `i2i.hd`, `qwen-image`, any other image channel, and `list_models` to pick a different channel. A runtime physical binding (e.g. `nano-banana2-max` or `qwen-image`) is acceptable only if the call passed `i2i.hd.lite` (has refs) or `t2i.hd.lite` (no refs). Wrong channel: discard and retry with the correct channel (protocol retry; remake gate not required).
- **Image-review remake gate**: After inspecting a generated image, if about to remake / retry / logic-redraw, do **not** remake in the same turn. Tell the user what fails, then `set_guidance` with **exactly two** buttons (translate labels), and **stop and wait**:
- `{"text": "Regenerate"}`
- `{"text": "Proceed to next step"}`
Do not mix with style, winner, language, or other buttons. Do not show one button or auto-retry. **Regenerate** (after user selects): remake with the same channel lock. **Proceed to next step** (after user selects): keep the current image and continue, even if imperfect. After a remake returns, inspect again; if about to remake again, show the same two buttons and wait. Applies to St