Kling AI

View saved version →

---
name: kling-ai
description: Generate Kling images and videos in local Codex, including text-to-image, image-to-image, text-to-video, image-to-video, reusable Elements, and motion control. Also handles login, credits, uploads, feedback, and task status through the Global CLI.
---

# Kling AI

## Capabilities

| Request | Command | Guidance |
| --- | --- | --- |
| Generate an image from text | `text_to_image` | [Image generation](../kling-ai-generate-image/SKILL.md) |
| Edit or restyle a reference image | `image_to_image` | [Image generation](../kling-ai-generate-image/SKILL.md) |
| Generate a video from text | `text_to_video` | [Video generation](../kling-ai-generate-video/SKILL.md) |
| Animate an image / use first and tail frames | `image_to_video` | [Video generation](../kling-ai-generate-video/SKILL.md) |
| Create a reusable subject from images or a video | `element_create` | [Create an Element](references/elements-motion-feedback.md#reusable-elements) |
| List saved subjects and their IDs | `element_list` | [List Elements](references/elements-motion-feedback.md#reusable-elements) |
| Read a subject's full details and resources | `element_get` | [Inspect an Element](references/elements-motion-feedback.md#reusable-elements) |
| Update a subject's name, description, tags, or supported resources | `element_update` | [Update an Element](references/elements-motion-feedback.md#reusable-elements) |
| Delete a specified saved subject | `element_delete` | [Delete an Element](references/elements-motion-feedback.md#reusable-elements) |
| List saved motions and their IDs | `motion_library_list` | [Browse saved motions](references/elements-motion-feedback.md#motion-library-and-motion-control) |
| Generate a video by driving a subject image with a motion video or saved motion ID | `motion_control` | [Generate with motion control](references/elements-motion-feedback.md#motion-library-and-motion-control) |
| Upload a local asset and return its URL | `file_upload` | [Explicit upload](references/elements-motion-feedback.md#explicit-asset-upload) |
| Query a generation's status and results | `query_tasks` | [Task results](references/cli-contract.md#output) |
| Check membership and remaining credits | `account` | [Account](references/cli-contract.md#commands) |
| Log in with browser OAuth | `login` | [Login](references/cli-contract.md#runtime) |
| Log out on the server and clear local credentials | `logout` | [Logout](references/cli-contract.md#commands) |
| Read the current identity, available models, and parameter specifications | `who_am_i` | [Model discovery](references/cli-contract.md#commands) |
| List the server's available tools and input schemas | `tool_list` | [Tool discovery](references/cli-contract.md#commands) |
| Report a stuck task, billing anomaly, or unexpected result with user authorization | `feedback` | [Issue feedback](references/elements-motion-feedback.md#issue-feedback) |

To reuse a saved subject in generation, inspect it with `element_get`, then bind it using `--elements` and `<<<id>>>` in a compatible generation command; see [binding rules](references/elements-motion-feedback.md#bind-an-element-to-generation). Creating an Element alone does not generate an image or video.

For a capability question, briefly describe image generation, video generation, reusable subjects, and motion control. Do not run a paid command.

Read [CLI commands and output](references/cli-contract.md) before the first remote operation. For Elements, motion control, explicit uploads, or feedback, also read [capability rules](references/elements-motion-feedback.md). Route image creation to `kling-ai-generate-image` and video creation/motion transfer to `kling-ai-generate-video`.

## Runtime and account

- Require shell access, Node.js 18+, and npm. Use `npm exec --yes --registry=https://registry.npmjs.org --package=@klingai/cli-global@0.2.0 -- kling --version`; use exactly CLI 0.2.0 for this plugin release. npm may download and execute that package from the official registry; explain this dependency before first use and respect host execution approvals. Do not change the version or bypass a denied command. If unavailable, report the limitation; do not substitute another generator or register an MCP server.
- Start with `who_am_i`. If authentication is needed, launch `login --skill-name kling-ai-cli --skill-version 1.0.5`; the user completes browser OAuth. Then run `tool_list` and `who_am_i`. Never request, read, print, or log credentials, tokens, or cookies.
- Use only the Global package and its service configuration. To switch accounts, run `logout` → `login` → `who_am_i`; stop if logout/login fails. Sign-out alone ends after `logout`.

## Submit and retrieve

1. A clear request authorizes one paid generation per explicitly requested step. Resolve only material ambiguities; do not add redundant confirmations or unrequested paid steps.
2. Select a live model from `who_am_i`; use its declared parameters, defaults, enums, and inputs. Rebuild flags after changing models. Preserve user facts and supplied copy; do not invent unsupported settings or commercial claims.
3. For pasted or attached media, resolve its path and role; pass correctly quoted arguments to the CLI for auto-upload. Never evaluate user text as shell code. Do not replace a failed reference workflow with text-only generation.
4. Immediately before every paid command, run `account`. Stop on zero or known insufficient credits. Keep at most one non-terminal paid task per objective; submit each authorized step at most once.
5. Preserve the returned `generationId`. Poll that task at provider-permitted intervals until terminal state, user cancellation, or turn timeout; a status-only request queries once. Return status, consumed credits when supplied, and the selected `works[]` URL. Stable identity is task number plus work index and content type. Result URLs expire after 24 hours; use a returned watermark-free URL only when authorized. Claim visual inspection only after doing it.
6. For dependent steps, continue only after the preceding step succeeds and the selected work is unambiguous; otherwise terminate the chain. Before reusing older Kling media, refresh its task once and use the same work index's current URL. If unavailable, request a fresh attachment; do not query repeatedly or select a substitute.

Element operations are synchronous: they return persistent subject IDs, not generation IDs. Do not poll them with `query_tasks`. Uploading an asset or selecting an Element does not authorize generation.

## Failures

- Only an explicitly pre-billing argument rejection with no `generationId` permits one correction: the original authorization remains unused. Refresh the live schema, preserve intent, recheck `account`, and submit the corrected command once. Stop after one correction.
- Credits/membership, queue/rate limits, media errors, safety restrictions, provider errors, or ambiguous billing: report the diagnostic and stop. Do not switch models, alter media, bypass restrictions, or automatically resubmit. Never claim a refund or non-billing without provider confirmation.
- If the submission response is lost, query only a known task number. Without one, report unknown state; a replacement requires that the user explicitly accepts possible duplicate billing. Feedback sends an issue summary and related task IDs to Kling. Send only with user authorization covering that data; existing authorization does not need repeated confirmation. State briefly when a report is sent and preserve host approval controls. Feedback does not fix or retry the task.

Referenced files: 3

View saved version →

---
name: kling-ai-generate-image
description: Generate or edit Kling images through the Global CLI, including references, product images, portraits, posters, variants, and reusable Elements.
---

# Kling image generation

Follow [the core workflow](../kling-ai/SKILL.md) and [CLI contract](../kling-ai/references/cli-contract.md); submit each authorized step at most once. Read [Element rules](../kling-ai/references/elements-motion-feedback.md) only when using reusable subjects.

- Use `text_to_image` for a new image; use `image_to_image` when source media controls identity, structure, composition, or style. Assign each reference a role and preserve the order of repeatable `--image` arguments.
- For editing, state what changes and what stays fixed. Preserve identity, product geometry, logos, and exact supplied copy. Write concrete composition, lighting, and material choices; avoid unnecessary prompt embellishment.
- Choose a live `--model`; use `--omni` only when explicitly requested and supported. Preserve the model's declared resolution default. Use only live-supported ratio and count values; default to one image unless more were requested.
- For Element binding, inspect the subject type and both live tool restrictions and model parameters. Use `<<<id>>>` plus `--elements`; stop if declarations conflict. Do not probe compatibility with a paid task.
- Resolve references, check credits, submit, poll, and return results using the core workflow. Status requests use `query_tasks` and create no generation.

## Quality workflow

Before every image generation, read [prompt construction](references/prompt-construction.md). For product, advertising, portrait, cover, or reference-edit requests, also read the relevant section of [scene patterns](references/scene-patterns.md). User facts and requested changes take precedence over these profiles.

1. Identify the subject, destination, required copy, reference roles, and protected identity/product facts. Ask only for missing information that materially changes the result.
2. Select one primary quality profile: product/ecommerce, advertising/poster, portrait/fashion, cinematic concept, or reference edit. Add another only when the request crosses contexts; profiles guide visible quality, not model cost.
3. Build one coherent prompt covering the subject and allowed change, composition, lighting, color, and material response. Translate “premium” or “cinematic” into observable choices. Keep model, ratio, resolution, and count in flags.
4. For references, separate identity/product structure from style/composition. Preserve source perspective, geometry, labels, and supplied text; define which areas may change. Do not redescribe a source so broadly that locked facts drift.
5. Before submission, check focal hierarchy, crop/copy-safe space, reference order, protected details, exact spelling, and the chosen profile's acceptance criteria. Prefer clean artwork with reserved text space unless generated copy is required; include required copy exactly once.
6. Inspect the result when available before claiming visual QA. Report visible mismatches honestly; a quality defect does not authorize automatic regeneration.

## Minimal calls

These show arguments after the npm command prefix in the CLI contract. Replace `<model>` with the selected live model; resolve real paths before calling. Append the shared telemetry flags.

```bash
kling text_to_image --model <model> "A red panda in a vintage spacesuit, lit by Earth's blue glow"
kling image_to_image --model <model> --image ./reference.png "Restyle in watercolor; preserve the subject and composition"
```

For multiple references, repeat `--image` in the intended order. Add `--imageCount`, `--imgResolution`, or `--aspectRatio` only when declared by the selected model and required by the request; preserve defaults otherwise. Choose a product's factual reference over a style reference for identity and structure.

Before execution: resolve references → `who_am_i` → choose model and flags → `account` → submit once → `query_tasks <generationId>` → return selected `works[]` with task number and work index. An upload failure does not authorize text-to-image fallback.

Referenced files: 3

View saved version →

---
name: kling-ai-generate-video
description: Generate Kling videos through the Global CLI from text, images, reusable Elements, or a motion video/saved motion; supports live audio and multi-shot options.
---

# Kling video generation

Follow [the core workflow](../kling-ai/SKILL.md) and [CLI contract](../kling-ai/references/cli-contract.md); submit each authorized step at most once. For reusable subjects or motion transfer, read [Element and motion rules](../kling-ai/references/elements-motion-feedback.md).

- Use `text_to_video` without controlling source media; `image_to_video` for image-based animation; `motion_control` for a subject driven by a video or saved motion. A single image with “animate this” normally supplies the first frame. Clarify roles only when materially ambiguous.
- Pass image references with `--image` and a live-supported tail frame with `--tailImage`. Motion control needs a subject image and exactly one `--video` or `--motionId`; it always requires `--model`, never `--omni`.
- Write the opening state, subject action, camera movement, and ending. Preserve identity/product facts. Keep shots within the supported duration; do not submit a task per shot unless separate outputs were requested.
- Preserve the model's declared resolution default. Infer ratio/duration only from supported values and user intent. Omit it when undeclared; never reuse another model's flags.
- When declared, map `prefer_multi_shots` to the planned shot structure. Set `enable_audio` true only for requested sound and otherwise false; use `enable_asmr` only for requested ASMR. Pass `audio_prompt` and `music_prompt` only when declared and requested. For motion control, source-motion constraints take precedence over ordinary video defaults.
- Element binding requires both `<<<id>>>` and live-supported `--elements`; check tool-description restrictions as well as model parameters. Conflicting declarations stop submission.
- Resolve references, check credits, submit, poll, and return results using the core workflow. Do not invent dialogue, claims, or an extra paid generation step.

## Quality workflow

Before every video generation, read [motion and shot planning](references/motion-and-shots.md). For product, UGC, explainer, multi-shot, or social-format requests, also read the relevant section of [scene patterns](references/scene-patterns.md). A simple single-shot animation does not need every scenario guide.

1. Establish opening/ending states, duration, destination, reference roles, protected identity/product facts, and requested sound. Choose a live model that can support these requirements.
2. Write in time order. Separate subject action, camera motion, and environmental response; describe direction, speed, amplitude, and an observable endpoint. Default to one primary action and one camera path for a short shot.
3. For image animation, treat the first frame as the factual baseline. State what stays fixed and what moves. For motion control, preserve source-motion intent; do not invent conflicting choreography or camera moves.
4. Use multiple shots only when requested or needed for changes in place, time, scale, or information. Give each shot one purpose and a duration; total timing must close. Preserve faces, clothing, product geometry, lighting, screen direction, and location anchors across cuts.
5. Map dialogue, narration, music, ambience, or ASMR only when requested and supported. Never claim a completed voiced video when the model lacks the required sound controls, or invent testimonials, product claims, or copy.
6. Before submission, check that actions fit the duration, camera moves do not conflict, the ending is achievable, and profile-specific continuity checks pass. Inspect output before claiming visual QA; report defects without automatically submitting another paid task.

## Minimal calls

These show arguments after the npm command prefix in the CLI contract. Replace `<model>` with the selected live model and use actual media paths. Append the shared telemetry flags.

```bash
kling text_to_video --model <model> "Sunset over the sea; gentle waves, slow forward camera movement"
kling image_to_video --model <model> --image ./first-frame.png "Animate the waves; preserve the horizon and composition"
```

For a requested ending frame, add `--tailImage ./last-frame.png` only when supported. Add `--duration` or `--aspectRatio` only with live-supported values. Describe multi-shot timing in the prompt and use declared shot/audio flags; a storyboard does not imply one paid task per shot. Motion transfer uses the separate `motion_control` workflow, not ordinary image animation.

Before execution: resolve media roles → `who_am_i` → choose model and flags → `account` → submit once → `query_tasks <generationId>` → return selected `works[]` with task number and work index. A media failure does not authorize text-to-video fallback.

Referenced files: 3