Kling AI Video Generation
Translate a user brief into a coherent Kling motion plan and one approved remote generation request. Use only live tools and schemas from the configured MCP at https://kling.ai/mcp.
Contract
- Use host-managed OAuth. Never request or expose API keys, tokens, cookies, authorization headers, or signed URLs.
- A user request to generate authorizes one submission after materially missing inputs are resolved. Do not add a credit-cost warning or a separate confirmation step.
- Submit once per approved intent. Never automatically retry a failed or ambiguous generation.
- Discover live tools and schemas at runtime. Do not hard-code model names, input roles, duration values, or multi-shot fields from examples.
- Upload attached media through the remote upload tool when required and preserve the exact returned reference.
Before submission, read the complete MCP input/output and current model parameter snapshot, then let the current tools/list and who_am_i override dynamic snapshot values.
Workflow
- Classify the request using the mode table below.
- Read scene patterns for the matching format.
- Read motion and shot planning for camera choreography, image-to-video constraints, multi-shot continuity, or timed narration.
- Ask only for missing creative facts that materially change the result: duration, destination ratio, required references, narration/copy, or shot structure.
- Among live models compatible with the mode, references, and required duration, prefer a full-quality model. Prefer a fast, Turbo, or low-cost model only when the user explicitly asks for a draft, speed, or credit savings.
- Build a motion-first prompt describing subject action, camera action, environmental motion, timing, continuity, and protected elements. Translate abstract requests such as “cinematic,” “premium,” or “high quality” into camera movement, lighting, materials, depth of field, motion rhythm, and composition instead of stacking adjectives.
- Call the selected live generation tool once when the request has enough information. Preserve
generationId and any taskTraceId.
- If the submission is not terminal, poll its status at provider-allowed intervals until success or failure. On user cancellation or current-turn timeout, return the current state and task number.
- Provide the primary video or result link returned by Kling. Show
generationId as the task number and keep taskTraceId internal unless troubleshooting requires it.
Generation modes
| User intent |
Mode |
Required interpretation |
| Text-to-video |
Generate |
No source image controls the opening frame. Define the opening composition from text. |
| Image-to-video |
Image-to-video |
One or more images control the first frame, last frame, identity/product reference, or visual reference. Assign each role explicitly. |
| Motion control |
Motion transfer |
Require a subject image and exactly one motion source: library motionId or a motion-source video. Obtain all other fields from the live model schema. |
| Storyboard |
Single approved video plan |
Split timing and continuity deliberately; do not submit one task per shot unless the user explicitly approves separate tasks. |
| Status check |
Read-only |
Do not call a generation tool; query the existing task. |
For image-to-video, distinguish these roles before submission:
- first frame: lock opening composition and animate forward from it;
- last frame: define the intended destination only when the live schema supports it;
- identity/product reference: preserve subject facts without assuming the input is the first frame;
- style reference: transfer only named visual traits, not identity or composition.
Do not silently fall back from image-to-video to text-to-video when upload, reference count, or schema validation fails. Report the limitation and let the user revise the request.
Before calling the tool, check the selected mode, reference roles, duration,
resolution, shot structure, and protected elements internally. Do not show a
pre-submission process message unless you need the user to clarify a missing
creative requirement.
Quality-first defaults
- Use
1080p for a normal deliverable, 4k when supported for high-quality, commercial, large-screen, or post-production work, and 720p only for drafts, speed/cost-first work, or a mode with no higher setting. Do not lower a higher live model default.
- Use
5 seconds for one action or one shot; prefer 10 seconds for dialogue, singing, a complete product action, or two connected beats; use a longer supported duration only when the narrative requires it. Choose the shortest duration that can complete the idea.
- Infer text-to-video ratio from the destination:
9:16 for vertical shorts, 1:1 for square feeds, and 16:9 for landscape ads, web, or YouTube. Use 16:9 only when no destination context exists.
- For image-to-video, derive composition from the source and avoid passing a ratio unless required.
- Prefer one continuous shot for a single moment. Use multi-shot only for explicit narrative progression, multiple locations/times, or a requested sequence.
- Keep the first generation focused. Do not add narration, on-screen copy, extra characters, or product claims that the user did not request.
Quality gate
Before submission, check that the subject action can fit the duration, camera instructions do not conflict, first/last frame intent is clear, reference identity/product geometry is protected, and multi-shot durations form a coherent whole. For ads and explainers, ensure every shot has a single communication job.
Failure behavior
- Authorization failure: direct the user to WorkBuddy's native MCP connection flow.
- Invalid argument/model: refresh the live schema and revise only the unsupported field.
- Insufficient credits: tell the user to recharge and stop. Do not retry automatically.
- Lost response: treat task creation as unknown and query existing tasks before any new submission.
- Provider failure: report the message and preserve IDs; never resubmit automatically.
1---2name: kling-ai-generate-video-23description: Generate cinematic-quality videos via Kling AI in WorkBuddy. Supports T2V & I2V with smooth motion and narrative coherence. Ideal for product demos, ads, short films, and social content.4---5
6# Kling AI Video Generation
7
8Translate a user brief into a coherent Kling motion plan and one approved remote generation request. Use only live tools and schemas from the configured MCP at `https://kling.ai/mcp`.
9
10## Contract
11
12- Use host-managed OAuth. Never request or expose API keys, tokens, cookies, authorization headers, or signed URLs.
13- A user request to generate authorizes one submission after materially missing inputs are resolved. Do not add a credit-cost warning or a separate confirmation step.
14- Submit once per approved intent. Never automatically retry a failed or ambiguous generation.
15- Discover live tools and schemas at runtime. Do not hard-code model names, input roles, duration values, or multi-shot fields from examples.
16- Upload attached media through the remote upload tool when required and preserve the exact returned reference.
17
18Before submission, read the [complete MCP input/output and current model parameter snapshot](../kling-ai-plugin/references/mcp-contract.md), then let the current `tools/list` and `who_am_i` override dynamic snapshot values.
19
20## Workflow
21
221. Classify the request using the mode table below.
232. Read [scene patterns](references/scene-patterns.md) for the matching format.
243. Read [motion and shot planning](references/motion-and-shots.md) for camera choreography, image-to-video constraints, multi-shot continuity, or timed narration.
254. Ask only for missing creative facts that materially change the result: duration, destination ratio, required references, narration/copy, or shot structure.
265. Among live models compatible with the mode, references, and required duration, prefer a full-quality model. Prefer a fast, Turbo, or low-cost model only when the user explicitly asks for a draft, speed, or credit savings.
276. Build a motion-first prompt describing subject action, camera action, environmental motion, timing, continuity, and protected elements. Translate abstract requests such as “cinematic,” “premium,” or “high quality” into camera movement, lighting, materials, depth of field, motion rhythm, and composition instead of stacking adjectives.
287. Call the selected live generation tool once when the request has enough information. Preserve `generationId` and any `taskTraceId`.
298. If the submission is not terminal, poll its status at provider-allowed intervals until success or failure. On user cancellation or current-turn timeout, return the current state and task number.
309. Provide the primary video or result link returned by Kling. Show `generationId` as the **task number** and keep `taskTraceId` internal unless troubleshooting requires it.
31
32## Generation modes
33
34| User intent | Mode | Required interpretation |
35| --- | --- | --- |
36| Text-to-video | Generate | No source image controls the opening frame. Define the opening composition from text. |
37| Image-to-video | Image-to-video | One or more images control the first frame, last frame, identity/product reference, or visual reference. Assign each role explicitly. |
38| Motion control | Motion transfer | Require a subject image and exactly one motion source: library `motionId` or a motion-source video. Obtain all other fields from the live model schema. |
39| Storyboard | Single approved video plan | Split timing and continuity deliberately; do not submit one task per shot unless the user explicitly approves separate tasks. |
40| Status check | Read-only | Do not call a generation tool; query the existing task. |
41
42For image-to-video, distinguish these roles before submission:
43
44- **first frame:** lock opening composition and animate forward from it;
45- **last frame:** define the intended destination only when the live schema supports it;
46- **identity/product reference:** preserve subject facts without assuming the input is the first frame;
47- **style reference:** transfer only named visual traits, not identity or composition.
48
49Do not silently fall back from image-to-video to text-to-video when upload, reference count, or schema validation fails. Report the limitation and let the user revise the request.
50
51Before calling the tool, check the selected mode, reference roles, duration,
52resolution, shot structure, and protected elements internally. Do not show a
53pre-submission process message unless you need the user to clarify a missing
54creative requirement.
55
56## Quality-first defaults
57
58- Use `1080p` for a normal deliverable, `4k` when supported for high-quality, commercial, large-screen, or post-production work, and `720p` only for drafts, speed/cost-first work, or a mode with no higher setting. Do not lower a higher live model default.
59- Use `5` seconds for one action or one shot; prefer `10` seconds for dialogue, singing, a complete product action, or two connected beats; use a longer supported duration only when the narrative requires it. Choose the shortest duration that can complete the idea.
60- Infer text-to-video ratio from the destination: `9:16` for vertical shorts, `1:1` for square feeds, and `16:9` for landscape ads, web, or YouTube. Use `16:9` only when no destination context exists.
61- For image-to-video, derive composition from the source and avoid passing a ratio unless required.
62- Prefer one continuous shot for a single moment. Use multi-shot only for explicit narrative progression, multiple locations/times, or a requested sequence.
63- Keep the first generation focused. Do not add narration, on-screen copy, extra characters, or product claims that the user did not request.
64
65## Quality gate
66
67Before submission, check that the subject action can fit the duration, camera instructions do not conflict, first/last frame intent is clear, reference identity/product geometry is protected, and multi-shot durations form a coherent whole. For ads and explainers, ensure every shot has a single communication job.
68
69## Failure behavior
70
71- Authorization failure: direct the user to WorkBuddy's native MCP connection flow.
72- Invalid argument/model: refresh the live schema and revise only the unsupported field.
73- Insufficient credits: tell the user to recharge and stop. Do not retry automatically.
74- Lost response: treat task creation as unknown and query existing tasks before any new submission.
75- Provider failure: report the message and preserve IDs; never resubmit automatically.
ahang1598