Skill instructions
kling-ai4.05 KB
View saved version →
---
name: kling-ai
description: Plan Kling AI image prompts and video storyboards, prepare or submit authorized Kling generations, and manage Elements, saved motions, credits, known tasks, and account access.
---
# Kling AI
## Route the request
- For prompt-only work, return the requested prompt or storyboard. For a missing required reference, ask for the asset and explain its role. Neither needs an account query.
- For `text_to_image` or `image_to_image`, use the [image Skill](../kling-ai-generate-image/SKILL.md). For `text_to_video`, `image_to_video`, `omni_ref_video`, or `motion_control`, use the [video Skill](../kling-ai-generate-video/SKILL.md).
- For Elements, browsing saved motions, or authorized feedback, use [the relevant tools](references/elements-motion-feedback.md). Resolve real IDs before use. An attachment or library operation does not authorize a generation.
- Check credits through the selected runtime. Model availability comes from `who_am_i`. A direct status request runs `query_tasks` once for the known `generationId`.
## Connection and account
Read [host routing](references/host-routing.md) to select the runtime before any tool call, account check, or login. Keep one backend, account, and region per objective. Use logout or switch accounts only on request. If authorization is required, explain the return-to-client step and let the original login process finish. Only report login complete after `who_am_i` succeeds with an authenticated identity on that same connection.
## Prepare and submit
1. Follow [model discovery and request validation](references/generation-settings.md). Prefer directly attached images and map each reference to its live input slot. Preserve one UUIDv7 `taskTraceId` for the objective where supported. Retain the returned `generationId` as the task number.
2. Show the final tool and model, prompt summary, output count, resolution, duration and ratio as applicable, reference roles, and requested audio, HDR or shot settings. Explain live billing rules, including per-second billing and input video duration when declared. A proposal ends here. Before generation, get explicit authorization for the final settings, reusing approval already given for these settings.
3. Verify the actual request against the live schema. Immediately before submission, check the same account's balance. Pause if verification fails, the balance is nonpositive, or the service confirms it is insufficient for these settings. A positive balance without a live estimate provides no cost guarantee. Each authorization covers at most one submission attempt. Keep at most one non-terminal paid task for a dependent workflow.
4. Preserve the returned task number and query only that task. If submission is ambiguous or the CLI exits with code 3, query a known task ID or report that creation and billing are unconfirmed. A trace ID or new balance cannot resolve an unknown submission. An explicitly rejected attempt requires fresh authorization before retry. A failed preview never authorizes generation.
Read [task recovery](references/task-recovery.md) for insufficient credits, continuation, missing results, or an expired link. For a request to purchase credits or change plans, explain the account limitation and keep generation paused. Do not initiate purchases or provide transaction links. Continue dependent tasks only after their prerequisite succeeds and the intended work can be identified.
## Results
Follow the selected runtime's [media and result rules](references/host-routing.md#media-and-result-boundaries). Give the task number and current status when helpful. Avoid parallel polling and duplicate media. If responsible for refresh, query only the original task within the [task recovery limits](references/task-recovery.md).
Use the task number plus work index when reusing or reporting a result. Check task and work-level status, media type, and actual usable count. A video cover is preview art. Deliver available works and re
Preview of the first 4,000 characters. Read full instructions →
Referenced files: 6 — file references are metadata, not package contents.
kling-ai-generate-image3.24 KB
View saved version →
---
name: kling-ai-generate-image
description: Write or refine Kling image prompts, or prepare and generate approved images and edits, including product photography, portraits, posters, variants, Elements, and continuous image series.
---
# Kling image generation
For prompt-only work, use [prompt construction](references/prompt-construction.md) and return the requested prompt. Ask for a missing required image before model discovery. Follow the [core workflow](../kling-ai/SKILL.md) for preparation, billing and results. For local calls use the [CLI contract](../kling-ai/references/cli-contract.md). Submit each authorized task at most once.
## Choose the mode and model
- Use `text_to_image` for a new image or series. Use `image_to_image` when source media controls editable content, identity, product geometry, composition, or style. Preserve each reference's role, slot and order.
- When model selection is open, prefer Kling Image 4.0 if the live tool lists it and it fits the request. Otherwise prefer a compatible live `defaultModel`, then another compatible model advertised for the same tool. Honor a specific model or a request to use the service default.
- A continuous image series needs a model that supports series generation at the requested count and resolution in one task. For Kling Image 4.0, use `story_mode` when declared and set `imageCount` when the selected mode honors it. If the model infers the series length from the prompt, state that the count is approximate and get approval before relaxing an exact count. Check that 8K, when requested, is available with the selected mode and count. Multiple independent images cannot fulfill a continuous series.
- For Elements, use [binding rules](../kling-ai/references/elements-motion-feedback.md#bind-an-element-to-generation). Verify type, required slots and combined reference limits using [live settings](../kling-ai/references/generation-settings.md).
## Quality workflow
1. Use [prompt construction](references/prompt-construction.md) to review the user's prompt. Read the relevant [scene pattern](references/scene-patterns.md) for product, advertising, portrait, cover, or reference editing when useful. Use a suitable supplied prompt verbatim.
2. State the visible subject, composition, lighting, and material. For editing, identify the requested changes and protect source identity, product geometry, logos, supplied copy and protected areas. Reserve copy-safe space when called for.
3. Preserve the live resolution default unless the user requests another value or approves a recommendation. Infer a supported ratio from the destination and choose one result by default. Check the exact wording of on-image text and the reference roles.
4. Follow the core workflow to check balance, submit once and inspect the media when available before claiming visual QA. An unavailable reference pauses the original request. Use the edit task's own ID to recover its output.
## Local CLI examples
Replace `<model>` with the canonical ID and use the verified executable from the CLI contract.
```bash
kling-plugin text_to_image --model <model> "A red panda in a vintage spacesuit, lit by Earth's blue glow"
kling-plugin image_to_image --model <model> --image ./reference.png "Restyle in watercolor; preserve subject and composition"
```
Referenced files: 3 — file references are metadata, not package contents.
kling-ai-generate-video3.68 KB
View saved version →
---
name: kling-ai-generate-video
description: Write Kling video prompts and storyboards, or prepare and generate approved clips with text, frames, Omni Reference, HDR, audio, and motion transfer.
---
# Kling video generation
For prompt-only or storyboard work, use [motion and shot planning](references/motion-and-shots.md) and return the requested plan. Ask for missing assets only when the intended setup requires them. Follow the [core workflow](../kling-ai/SKILL.md) for preparation, billing and results. For local calls use the [CLI contract](../kling-ai/references/cli-contract.md). Submit each authorized task at most once.
## Choose the current mode
- Check live `text_to_video` for text, `image_to_video` for frame-driven requests, and `omni_ref_video` for both when available. A text-only Omni Reference pair must allow no media. A frame-driven pair must declare the requested frame roles. For mixed media, check live support for every role and their combinations.
- With open model choice, prefer compatible Kling 4.0 Omni Reference when the account offers it. For speed or credit savings, consider a compatible Kling 4.0 Flash pair. Honor an explicitly named 4.0 or 4.0 Flash variant. Match it to the current full ID or declared alias and validate the tool/model pair.
- Standard 4.0 can offer first, tail and intermediate keyframes. Flash image-to-video uses a single first frame where declared. On CLI, when using named frame slots, bind them with repeatable `--input` such as `--input first_image=./first.png`. Use either named frame inputs or `--image` / `--tailImage` shortcuts for that command.
- For motion transfer, use `motion_control` with a subject image and exactly one motion source, following [Element and motion rules](../kling-ai/references/elements-motion-feedback.md). For HDR, required silence, multiple reference types, or source audio retention, use [live settings](../kling-ai/references/generation-settings.md) to filter eligible pairs before selecting one.
## Quality workflow
1. Use [motion and shot planning](references/motion-and-shots.md) and the relevant [scene pattern](references/scene-patterns.md) for the destination. Preserve the characters, product facts and continuity that anchor the scene.
2. Write in time order. Describe the opening, subject action, camera movement and ending separately. Use one primary camera movement for a short shot. For a multi-shot request, keep shot timings within total duration and submit one task. A shot preference expresses intent. Require explicit support when the user needs a fixed shot count or a single-shot guarantee.
3. Honor requested ratio, duration and resolution when supported. With unspecified ratio, preserve first-frame composition when the model permits it. Suggest about 5 seconds for one beat or 10 seconds for two connected beats when supported. Keep live defaults for unspecified settings unless the user approves a change.
4. Preserve exact dialogue and synchronize audio cues with the visible events when requested. Verify supported audio controls. Do not add narration, readable text or product claims without a request.
5. Follow the core workflow to disclose applicable billing, check balance, submit once and inspect the actual video and audio before claiming visual QA. Use the video task number for recovery and reference task numbers only to refresh their source media.
## Local CLI examples
Replace `<model>` with the canonical ID and use the verified executable from the CLI contract.
```bash
kling-plugin text_to_video --model <model> "Sunset over the sea; gentle waves, slow forward camera movement"
kling-plugin image_to_video --model <model> --input first_image=./first-frame.png "Animate the waves; preserve the horizon and composition"
```
Referenced files: 3 — file references are metadata, not package contents.