AI 视频提示词优化 by CreatOK

概述

optimize-prompt

This skill optimizes a raw prompt for CreatOK AI video generation. It does not submit a video generation task by itself.

Constraints

• Platform: TikTok-style AI video prompts.
• The model's final user-facing response should match the user's input language, default English.
• Use creatok prompt-optimize; never call HTTP endpoints directly.
• Prompt optimization is synchronous and does not deduct user credits.
• Support up to three reference images when allowed by both capabilities.prompt_optimization.max_reference_images and the selected model's limits.max_reference_images.
• If the user has not provided the raw prompt to optimize, ask for it directly before running any CLI command.
• In normal user-facing replies, do not explain command names, JSON options, output paths, doctor/capabilities checks, or how the request is submitted. Keep those details internal unless the user explicitly asks for debugging details.
• When asking for a missing raw prompt, if a default video model has been resolved internally, do not present it as a setup detail. Instead, add a short user-facing note that the default model will be used unless the user wants another video generation model.
• After receiving the optimized prompt, show it to the user and proactively ask whether they want to use it for video generation.
• If the user wants to use it, continue into the normal video generation flow with the optimized prompt. Do not auto-submit video generation; the video generation skill must still perform its own confirmation gate.
• Follow shared guidance in ./references/common-rules.md.

Internal CLI Usage

All API interaction goes through the CLI. Every command prints a single JSON envelope to stdout; human logs go to stderr.

Commands used by this skill:

creatok doctor
creatok capabilities
creatok assets create --type image --file <path>
creatok prompt-optimize --prompt "<prompt>" [--reference-image-keys <key-1,key-2,key-3>] (--options '<json>' | --options-file options.json) --out <run_dir>

The optimized prompt is returned at data.prompt. The CLI also writes result.json under artifacts.dir when --out is provided.

Do not show these commands or paths to the user during a normal optimization flow.

Parameters

• Resolve the default video model from creatok capabilities using capabilities.video.default_model.
• Pass the resolved default model through --options as model unless the user explicitly chose another supported model.
• Always pass locale in --options; use the requested output language or explicitly use "en".
• Do not pass duration; the current prompt-optimization API does not apply it to the optimized prompt.
• For local reference images, create one formal Asset per image, preserve their order, and pass the returned keys through --reference-image-keys.
• If the user provides one to three reference images as conversation attachments or accessible local files, use all of them when the selected model supports that count.
• If the user provides a specific image URL and the agent can access it, download it to an artifact directory and pass that local file as the reference image.
• If the image URL cannot be accessed, ask the user to upload the image in the conversation or provide a local file path.
• For existing formal image Asset keys, preserve their order and pass them through --reference-image-keys.
• Do not combine reference-image flags with referenceImages, referenceImageKey, or reference_image_key in --options.
• Use --options-file instead of inline JSON when shell quoting is unreliable, especially in PowerShell.

Missing Prompt Handling

If the user asks to optimize a prompt but does not provide the actual prompt text, ask for it directly and keep the wording user-facing. Do not mention setup checks or command execution. Do not say that the CLI is ready, the environment was loaded, doctor passed, skills have no drift, or that creatok prompt-optimize will be called.

Recommended Chinese shape:

请把需要优化的原始提示词发给我。

如果方便，也可以一起补充：
- 是否想指定其他视频生成模型；不指定的话我会使用默认模型
- 优化后的提示词语言
- 是否有参考图；有的话可以直接上传到对话框，或给我一个具体图片链接

Ask for optional context without making it sound mandatory. If the user only provides the raw prompt, proceed with defaults.

Workflow

1. Collect the raw prompt

• If no raw prompt is available in the current conversation, ask for it using the Missing Prompt Handling guidance and stop.
• If the prompt is available, collect optional context only when the user already provided it or when one short clarifying question would materially improve the result.
• Optional context includes target video model, desired prompt language, and up to three reference images.

1. Entry self-check

• Run creatok doctor.
• If the command is not found, the CLI is not installed; follow ./references/common-rules.md.
• If data.api_key_configured is false or error.kind is auth, guide the user to configure the channel-specific key: CREATOK_API_KEY for stable, CREATOK_API_KEY_PREVIEW for preview.
• Check for a newer CLI and, if the user agrees, update before proceeding.
• If data.skills_drift is true, run creatok skills install to resync.
• Do not summarize successful setup checks to the user unless the user is debugging.

1. Resolve defaults

• Run creatok capabilities.
• Use capabilities.video.default_model as the optimization model context.
• If that field is missing, stop and report that a supported target model could not be resolved; never omit model.
• Keep the resolved model internal unless the user asked which model will be used.

1. Optimize

• Build a strict --options object containing both required fields: model and locale.
• If there are local reference images, create one formal image Asset for each, keep every returned data.asset.object_key, and preserve source order.
• Run:

creatok prompt-optimize --prompt "<raw prompt>" [--reference-image-keys <key-1,key-2,key-3>] --options '<json>' --out <run_dir>

1. Respond

• Parse stdout. If ok:true, read data.prompt.
• Show the optimized prompt in the conversation.
• Ask whether the user wants to use the optimized prompt to generate a video.
• Do not include the CLI command, options JSON, object key, run directory, or artifact path in the normal response.
• If the user says yes, continue with creatok-generate-video using this optimized prompt as the prompt/brief. Do not run creatok video from this skill before the video generation confirmation gate.
• If error is present, map error.kind using ./references/common-rules.md and explain plainly.

Handoff

When handing off to video generation:

• Carry forward the optimized prompt verbatim.
• Carry forward the chosen model and every reference-image choice that remains valid, preserving reference-image order so [Image 1], [Image 2], and [Image 3] keep the same meaning.
• Tell the video generation flow that this prompt has already been optimized, so it should not ask the user to rewrite it again unless the user requests changes.