專為 Suno 與 Udio 設計的 AI 音樂提示詞（Prompt/Tags）生成助手。當使用者提出抽象的情緒、畫面或曲風融合需求（如：「古風結合賽博龐克」）時，請觸發此技能。它會將模糊的靈感精準轉譯為專業的音樂流派、樂器、混音與結構標籤，大幅提升 AI 音樂生成的品質與準確度。
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新
無料GitHub で公開
gpt-image-2
Generate or edit images with OpenAI's gpt-image-2 via GMI Cloud's async request-queue API (console.gmicloud.ai). One dual-mode script: with NO input image it runs `gpt-image-2-generate` (text-to-image); with one or more input images it runs `gpt-image-2-edit` (one request per image). Submits the job, polls until done, and downloads every result image to the shared project-root `outputs/` folder. Defaults: quality=medium, output_format=png, size=1920x1088, n=2. API key is read from a `.gmi_api_key` file in the skill folder (also accepts `--api-key` or `$GMI_API_KEY`). ACTIVATION IS BY EXPLICIT MODEL NAME ONLY. This skill ACTUALLY GENERATES / EDITS real images and spends API credits, so TRIGGER ONLY when the user explicitly names this model / provider — "gpt-image-2", "gpt image 2", "GMI", "GMI Cloud" — e.g. "用 gpt-image-2 生圖", "用 gpt-image-2 改圖", "GMI 生圖", "GMI Cloud 生圖", "跑 gpt-image-2", "gpt image 2 編輯", "gpt image 2 生成". DO NOT TRIGGER on generic image wording that does NOT name gpt-image-2 / GMI — e.g. "編輯圖片", "圖片編輯", "載入圖片編輯", "圖片編輯技能", "改圖", "修圖", "編輯這張圖", "轉為黑白線稿", "圖片生成", "生圖", "image editing", "edit image", "image generation". Those are PROMPT-WRITING requests, NOT generation: • editing an image / a hypothetical visual change → use `prompt-master-image-editor` (writes an English edit prompt only, never generates an actual image). • other prompt tasks → the relevant `prompt-master-*` skill. Also DO NOT TRIGGER for: Stable Diffusion / TensorArt generation (use `sd-image-gen`), Gemini / nano banana editing (use `nano-banana-pro-edit`), upscaling (use `sd-upscaler`), or image selection / judging (use `image-judge`).
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-courseスター 4(GitHub 4 + cccskills 0)
実行スコア 未検証
インストール方法を見る含まれるファイル（7）
- SKILL.md7.4 KB
- docs/examples_zh-TW.md3.4 KB
- docs/flags_zh-TW.md5.5 KB
- docs/SKILL_zh-TW.md6.8 KB
- references/examples.md3.3 KB
- references/flags.md5.3 KB
- scripts/gpt_image_2.py20.9 KB
SKILL.md（原文）
インストールする前に、エージェントに与えられる指示の中身を確認できます。
gpt-image-2
Run scripts/gpt_image_2.py to call GMI Cloud's gpt-image-2 models. The
script is dual-mode and auto-detecting:
- No input image → generate (
gpt-image-2-generate, text-to-image). - One or more input images → edit (
gpt-image-2-edit, one request per image).
Every result image is saved to the shared outputs/ folder at the project
root, prefixed gpt_. The drop-zone for edit sources is the shared uploads/
folder next to it. Both are defined once in .claude/lib/media_paths.py and
shared with the gemini-3-pro-image and seedance-2-0 skills:
<project root>/uploads/ drop edit-source images here
<project root>/uploads/done/ sources move here after a successful run
<project root>/outputs/ every result lands here (gpt_ / gemini_ / seedance_)
Core directives
- Never hard-code the API key. It is resolved in this order:
--api-keyflag →$GMI_API_KEY(incl. project.env) →.gmi_api_keyfile in the skill folder (this is where the copied key lives). If none is found the script exits with a clear message. The same GMI key powers thegemini-3-pro-imageskill. - Run from project root — CWD is
/Users/kouzuimac/Documents/反重力/0527課用教材/biz-ai-video-course. (The shared folders resolve from the script's own location, so another CWD still works; project root just keeps the relative paths below readable.) - Ask for the prompt. No default prompt. If the user doesn't supply one, ask before running.
- Mode is decided by whether an image is supplied. Pass image
path(s)/URL(s) as positional args (or
--image) → edit. Pass none → generate.--mode generate|editforces it. - Long prompts →
@file.txt. For prompts longer than a couple of lines, save to a file under project root and pass--prompt @prompts/x.txt. Avoids shell-quoting traps. - Defaults: quality=medium, png, size=1920x1088, n=2. Override via
--quality / --output-format / --size / --nonly when asked. - Size is flexible (verified live). gpt-image-2 accepts any
WxHwith edges multiple of 16, longest edge ≤ 3840, aspect ≤ 3:1, total pixels 0.65M–8.3M — so up to 4K (3840x2160). Default1920x1088(native 16:9 HD). Exact1920x1080is rejected (1080 isn't ÷16) — use1920x1088.autoalso works. - Don't chain into posting. Report saved paths; do not auto-invoke
multi-sns-post. - Default source folder — the shared
uploads/. Edit-source images live in the project-root drop-zoneuploads/, shared with the other two generation skills. When the user wants to edit an image but doesn't give a full path (or just pastes/drops one), look there first. The script resolves bare filenames and a directory path against this folder automatically (also CWD, project root, anduploads/done/), so you can pass just2026-06-10-23.png, or point at the wholeuploads/directory to edit every image inside it in one run (a directory ref lists only its top level, so already-archived images indone/are not re-edited). Match the output--sizeto the source aspect ratio (e.g. a 9:16 portrait →1088x1920) so the edit isn't stretched or cropped. - Used sources self-archive. After a successful edit, sources that were
sitting in
uploads/are moved touploads/done/, so the drop-zone only ever shows what still needs doing. A failed edit leaves its source in place for a retry. Pass--keep-refswhen the user wants several prompts on the same image; a re-run still accepts the bare filename either way.
Knowledge hub
Read only the reference file relevant to the task:
- Need CLI flags, modes, exit codes, the GMI API contract (endpoints, payload fields, polling, response shape), output naming → references/flags.md
- Need concrete invocation examples (generate, edit single/many, file-based prompt, URL input, overrides, stdout shape) → references/examples.md
Standard operating procedure
- Get the prompt from the user (inline or
@file). - Decide mode: did the user provide an image — a path, or one sitting in the
shared
uploads/drop-zone? → edit, else generate. For edits without an explicit path, checkuploads/first. - Run the script (see
references/examples.mdfor forms):# generate (no image) python3 .claude/skills/gpt-image-2/scripts/gpt_image_2.py \ --prompt "<text-or-@file>" # edit (one or more images) python3 .claude/skills/gpt-image-2/scripts/gpt_image_2.py \ --prompt "<text-or-@file>" path/to/source.png - Watch stdout:
[mode] ...,[submit] ...,[poll] status=...,[✓] ...,📦 ...(source archived),[done] .... - Report saved
outputs/paths to the user.
Output protocol
- Files land in the shared project-root
outputs/folder (auto-created):- generate:
gpt_{YYYYmmdd_HHMMSS}[_k].{png|jpg} - edit:
gpt_{YYYYmmdd_HHMMSS}_{source_stem}[_k].{png|jpg} _kindex suffix is added only whenn > 1.- The
gpt_prefix is what tells these apart fromgemini_/seedance_results in the same folder. An existing file is never overwritten — a colliding name gets a_1,_2… suffix.
- generate:
- Stdout prefixes:
[mode] / [submit] / [poll] / [✓] / [batch] / [done] / [!]. Parse[✓]lines for paths when chaining into other tooling. - Exit codes:
0all good ·1config/validation error (bad key, invalid size/quality/format/n) ·2job ran but ≥1 image failed.
レビュー
まだレビューはありません。使ってみた感想をお寄せください。
コメント
ログインなしで書き込めます。使ってみた感想、質問、動かなかった環境などを気軽にどうぞ。
まだコメントはありません。最初のコメントを書いてみませんか。
同じリポジトリのスキル
AI Short-Drama Director & Clip Prompt Architect (AI 短劇導演 × 分鏡 Clip 架構師). Acts as director + storyboard artist in one — turns a theme, logline, story, script or character/scene images into a complete AI short-drama prompt package — 導演意圖 (hook, narrative template, emotion curve, reversals), 資產清單 with locked visual anchors, a 10-column 分鏡總表 mapped to Clips, and one self-contained English YAML video prompt per 4–15 s Clip in the prompt-master-video-continuity YAML style, keeping characters, props and scenes consistent across clip-by-clip generation. Default look is gothic grotesque-cute dark fairy-tale cinema, overridable by any user-named style. Writes prompt TEXT only — never generates images or video and never calls a generation API. Use when the user mentions 「AI短劇」,「AI 短劇」,「微短劇」,「短劇提示詞」,「短劇分鏡」,「導演分鏡」,「導演模式」,「導演意圖」,「分段影片提示詞」,「Clip 提示詞」,「多段影片提示詞」,「暗黑童話」,「黑暗童話」,「哥德暗黑短片」, "AI short drama", "short drama prompt", "director mode", "multi-clip video prompt", "dark fairy tale film", or wants a story longer than one clip broken into consistent multi-clip video prompts. Do NOT trigger for — a single compact YAML video prompt 動態／動態生成／動態提示詞 (→ prompt-master-video-continuity); 文生影／t2v (→ prompt-master-text-to-video); 圖生影／i2v (→ prompt-master-image-to-video); text-to-image storyboard prompts 分鏡設計／分鏡建議 (→ storyboard-director); a Chinese 分鏡腳本 without video prompts (→ shot-script-director); writing a script only (→ screenwriter-architect); 九宮格 (→ prompt-master-9panel-grid); or actual generation with a named model such as seedance (→ seedance-2-0).
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新
Blocking Plan Director (運鏡平面圖導演). Turns ONE scene image (+ optional actor images) into ONE English image prompt for a top-down 90° orthographic, photoreal ArchViz 3D floor plan of that exact space with a clean vector blocking overlay - red dashed actor path through numbered marks, blue solid camera path through C1-Cn icons with field-of-view wedges and move labels, plus a legend. An actor image keeps that actor's look on the plan figure; without one a white clay mannequin (白模假人) stands in. Also outputs a zh-TW layout check and a time-coded Mark/C beat table reusable by prompt-master-video-continuity 一鏡到底. Generates immediately, prompt text only. Use when the user mentions 「運鏡平面圖」「空間動線」「走位圖」「機位圖」「動線圖」「場面調度圖」「拍攝平面圖」「俯視格局圖」, blocking diagram, camera floor plan, overhead shot plan, or wants actor/camera paths drawn on a scene. Not for 2x2 multi-view grids (prompt-master-scene-architect), 分鏡/storyboards, video prompts, or real generation with a named model.
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新
Hollywood Screenwriter & Character Architect (Oscar-grade, with a sociologist's and psychologist's acuity). Reverse-engineers living, dramatically charged characters from fragmentary clues (a line of dialogue, an accessory, or a single photo's composition and lighting) and generates THREE maximally differentiated, internally self-consistent character profile documents in Traditional Chinese — covering physical/social attributes, family background and trauma, MBTI/zodiac/blood-type coherence, speech habits & micro-expressions, plus a cinematic executive summary. Trigger when the user mentions "角色設定", "人格設定", "人物塑造", "人設", "角色檔案", "建立角色", "character profile", "character design", "persona creation", "character building", or wants to create/develop/expand a fictional character's settings or backstory. Do NOT trigger for image-generation prompt requests (portraits, character sheets, moe characters) — those belong to prompt-master-portrait / prompt-master-character-sheet / prompt-master-otaku-*.
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新
Chief Character Costume & Visual Design Director. Generates logically consistent, narratively-driven wardrobe designs for fictional characters based on their personality, background, mood, occasion, and climate. Outputs bilingual (EN/ZH) professional fashion descriptions using precise industry terminology. Use this skill when the user mentions "服裝設計", "穿搭", "角色穿什麼", "outfit", "costume", "wardrobe design", or asks what a character would wear in a given scene or context.
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新
Cultural Resonance Naming Master (文化共鳴命名大師) — a screenwriting-grade naming & renaming consultant that maps a character (or any narrative entity — faction, place, weapon, project, story title) onto a global mythological archetype and forges a name carrying metaphor, foreshadowing, and historical weight. Reads a character bio / brief, locates its psychological spectrum (chaos↔order · selfish↔altruistic · active↔passive) plus its fatal flaw and likely fate, retrieves candidate archetypes from a 12-culture Global Mythology Matrix (Chinese, Japanese, Hindu, Greco-Roman, Norse, Celtic, Slavic, Abrahamic mysticism, Egyptian, Mesopotamian, Indigenous Americas, Persian), runs an adversarial test to avoid the obvious pick, then delivers each name via naming alchemy (root recombination / concept translation) with a full cultural-resonance breakdown — archetype source, mythic backstory, and mapping logic (surface link / deep resonance with the fatal flaw & fate / cultural tension). Output analysis is in Traditional Chinese. Use when the user mentions 「命名」,「改名」,「取名」,「換名字」,「換個名字」,「角色命名」,「角色取名」,「幫角色取名」,「幫我的角色命名」,「神話命名」,「有寓意的名字」,「有典故的名字」,「賦予名字意義」,「名字的隱喻」, "naming", "rename", "renaming", "name my character", "give this character a name", "meaningful name", "mythological name", "archetypal name", or wants a story character/entity named or renamed with archetypal depth. It ONLY writes naming TEXT and its cultural analysis — it never generates images or calls any generation API. Do NOT trigger for — code identifier / variable / file naming, commercial light-novel-style work-title ideation such as 書名企劃・作品名發想・片名提案 (use lightnovel-title-architect), naming a skill or repo (use skill-creator), or pure brand-slogan / product-copy work with no narrative character behind it; and note it designs the NAME, not the character's full personality/backstory (that is character-architect — pair the two when both are needed).
日本語の概要は準備中です。原文の説明を表示しています。
Kouzui-HUANG/biz-ai-video-course42026年10月27日 更新