Adonis0123/adonis-skills / skills/prompt-studio
prompt-studio
FlaggedTurn a one-line image or video request into a finished asset: route it to an image, a code-rendered Code Video or a Model Video, search many GitHub prompt libraries for proven image prompts, video-model prompts (Seedance, Kling, Veo, Hailuo/MiniMax, Wan, Grok Imagine) and code-video templates, show the best 5 with their samples, model, author and license, write a generation brief, hand it to an available channel (host tool, Codex/Grok CLI, API, Hyperframes), then check and log the result. Use whenever the user wants a cover, illustration, poster, thumbnail, infographic, animation, motion clip, AI video, or a better prompt for one, even if they only say 'make me an image'. Also for Chinese-language requests for WeChat covers, illustrations, still-to-motion clips, and image/video prompt optimization. Not for general chatbot or coding prompts, photo editing without generation, UI implementation, or slide decks.
references/model-video.md
View on GitHub# Model Video
A video model generates the frames. Realistic, not exactly controllable, billed per second. Rules below were checked on 2026-10-27; models change fast, so re-check a model's official page when a run behaves unexpectedly and mark unchecked details `UNVERIFIED`.
## Before every run
This is a hard cost gate, and it covers the whole path: first-frame images made for image-to-video count too, so nothing paid runs before the yes. State each step (first frame, preview clip, final clip) with channel, model, duration, resolution, number of clips and estimated cost, and wait for a yes. A subscription channel (Grok Build, Codex) spends the user's plan quota for the media, and its CLI also reports an agent-call cost (`total_cost_usd`, about $0.05 per call in tests): list both, mark any price you could not check as unverified rather than guessing, and never say nothing will be charged. Run exactly one short low-resolution preview before the final, and ask again before the final run. Ambiguous 动效 / 动起来 requests do not come here: SKILL.md routes them to Code Video unless the user wants realistic motion or names a video model.
## Prompt structure (shared by all models)
1. Subject → action → setting → camera → light and style → sound/dialogue. Duration, aspect ratio and resolution are parameters, not prose.
2. One camera move and one main action per shot. Write motion as from-to ("walks from the doorway to the window, stops, turns"); a bare verb makes the subject wobble in place.
3. Repeat fixed traits (outfit, colours, 3–5 hex values) in every clip of a sequence.
4. Describe what should be there instead of listing what should not.
5. Image-to-video: the image fixes the first frame; the prompt only adds what the image lacks — motion, camera, sound. Do not re-describe the picture.
6. Product shots: say the product does not change ("the bottle's shape, colour and logo stay exactly as in the image"). Write light as something that happens to the surface ("a soft reflection slides across the glass"), never as a thing ("a band of light"): models draw a named thing as a new part of the product. Library ad scripts with several beats get cut to one beat: keep their camera move and light, drop their story.
## Per model
| Model / Channel | Duration | Output | References | Prompt specifics |
| -------------------------------------- | --------------------------------------- | ---------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Grok `reference_to_video` | 1–15 s (6 s recommended) | 480p or 720p | `images` up to 14 (re-rendered, referenced as `<IMAGE_i>`), `first_frame` / `last_frame` pinned exactly, up to 4 `keyframes` on 1/3 s steps, up to 3 preset `voices` (`<AUDIO_0>`) | No text-only video: at least one reference is required. Labels are numbered in upload order: first_frame, images, keyframes, last_frame. Named real people only from a user-supplied photo |
| MiniMax Hailuo 2.3 / 02 | 6 s or 10 s (1080P: 6 s only) | 768P default, 1080P | First and last frame: Hailuo-02 only; resolution follows the first frame | Camera commands in brackets, executed together when in one bracket (max 3): `[Truck left]` `[Truck right]` `[Pan left]` `[Pan right]` `[Push in]` `[Pull out]` `[Pedestal up]` `[Pedestal down]` `[Tilt up]` `[Tilt down]` `[Zoom in]` `[Zoom out]` `[Shake]` `[Tracking shot]` `[Static shot]`; sentence order is time order. Prompt ≤ 2000 chars. Set `prompt_optimizer: false` for exact control |
| MiniMax H3 / H3 Max | H3 4–15 s, Max 5–15 s | H3 768P or 2K; Max 480P or 768P | 0–2 images as first/last frame; reference mode up to 9 images, 3 videos, 3 audio (≤ 12 files) | Camera words are lowercase in brackets placed after the key description: `[pan]`, `[zoom]`, `[static]`. Prompt ≤ 7000 chars |
| Veo 3.1 (Vertex AI GA `veo-3.1-*-001`) | 4, 6 or 8 s (reference-image mode: 8 s) | 720p / 1080p / 4K (no 4K on Fast and Lite); 16:9 or 9:16; 24 fps | Text, image, first+last frame, extend | Generates audio; describe sound in separate sentences. Dialogue: `the man in the red hat says: Where is the rabbit?`. English prompts. Exclude things by naming them in the negative field, without "no"/"don't". The Gemini API preview models were shut down on 2026-10-22 |
| Seedance 2.x | `UNVERIFIED` (community: 4–15 s) | `UNVERIFIED` | `UNVERIFIED` (third parties: up to 9 images + 3 videos + 3 audio) | Community pattern: one opening line for format/style/palette, then time-coded shots `[00:00-00:05]` each with shot size + camera, then action and micro-expression; dialogue in quotes with tone in brackets; declare reference roles (`@Image1 is the character master`). Libraries: `recommend --kind video --model seedance` (LearnPrompt, YouMind, ai-boost) |
| Kling 3.0 | `UNVERIFIED` | Native 4K (official blog) | Element references, multi-shot with per-shot duration and camera | Clear scene direction; API camera control on 6 axes from -10 to 10 |
| Sora 2 | — | — | — | API shut down on 2026-09-24. Its prompting lessons still apply: shorter prompts leave the model more freedom; two 4 s clips often beat one 8 s clip |
Price guide (official pages, 2026-10-27): Hailuo 2.3 768P 6 s $0.28, 10 s $0.56, 1080P 6 s $0.49; Hailuo 2.3-Fast 768P 6 s $0.19; H3 2K $0.13/s, 768P $0.08/s. Other models: check the provider before quoting.
## Image to video, and sequences
1. A text-only request on a channel that needs an image (Grok `reference_to_video` needs at least one) starts with a first frame: generate it with the best available image Channel (Codex `image_gen` for products and photoreal, Grok `image_gen` when Codex is absent) at the video's aspect ratio, look at it, and count it in the cost gate. When the user supplied the image, use it as the first frame directly. Make the first frame with an image Channel first; when several characters or props share a shot, compose them into one image rather than pasting cut-outs (pasted cut-outs animate like stickers). Keep passing each character's reference sheet.
2. A subject crossing the frame: first frame at the start position, last frame at the end. Same image as first and last frame gives a seamless loop. Look at a staged last frame before spending on the video: an edit that moves the subject often moves other things too, and the video then animates that drift.
A busy first frame (many props, tiles or fine detail) warps and duplicates elements when animated. Stage a simpler frame, or keep everything but one subject still and say so explicitly.
3. Continuous shots: take the real last frame of the previous clip as the next first frame — `ffmpeg -sseof -0.05 -i shot1.mp4 -frames:v 1 last.png`. New camera angle: make a new first frame, same references.
4. Keep resolution and fps identical, then join without re-encoding: `ffmpeg -f concat -safe 0 -i list.txt -c copy out.mp4`.
## Quality gate
1. `studio.mjs check-video <file> --ratio W:H --duration S`: ratio within 1%, duration within 0.3 s. A `fail` is not delivered as is: apply its `fix` (pad, never stretch), check again and look at the result; say in one line when the channel's native size differs from the request (Grok 480p is 736×400, not exactly 16:9). `unverified` (no ffprobe) stays unverified in the report.
2. Extract one frame per second (`ffmpeg -i out.mp4 -vf fps=1 frames/%03d.png`) and look: identity drift, warped hands or faces, text artefacts, the motion actually happening.
3. On a moderation block: stop, tell the user, offer a different direction. Do not paraphrase to evade the filter.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |