Guiding MiniMax H3 Agents with Literal First/Last Frames

Where this fits

A user drops two images into the agent: "this is the first frame, this is the last frame, make the video." Without a specialized skill, an agent may treat both images as style references, causing the model to preserve unrelated details or ignore the exact endpoint composition. This skill is a narrow downstream specialist for MiniMax H3 that converts literal boundary frames into a continuous visual path—no more, no less.

Why agents benefit

• It routes only to explicitly declared first/last frame prompts, preventing the agent from overloading a keyframe with character or object consistency responsibilities.
• It generates temporal endpoint framing that anchors the video between two concrete images, rather than a vague style blend.
• It covers pure I2VA, FL2VA, and L2VA boundary-frame cases, so the agent can adapt the prompt shape to the input mode.
• It works cleanly after minimax-h3-creative-director has decided the images carry no reusable reference identity.
• It reduces prompt-repair loops by telling the video model exactly what remains variable: the motion between the declared endpoints.

Practical scenarios

Logo or title transition from locked start and end layouts

A designer provides a static logo on a white canvas as the first frame and the same logo scaled into a 3D scene as the last frame. The agent uses the keyframe skill to describe the motion path and camera movement without inventing a new logo style.

Time-of-day environment transition

A user supplies a dawn landscape and a night cityscape as endpoint frames. The agent builds a boundary-frame prompt that moves lighting, shadows, and atmospheric color from one image to the other, without assigning character or object references to either frame.

Product hero shot with exact composition endpoints

A marketing team gives a product hero image and a final frame with the product rotating and the background blurred. The skill lets the agent write a MiniMax H3 prompt that treats the second image as a destination composition to reach, not a texture reference to copy.

Add it to your agent workflow

Install it with the OpenAgentSkill CLI:

npx skills add unknowlei/minimax-h3-opencode-skills --skill minimax-h3-keyframe-video-prompt

After installation, call it only when the user explicitly labels an image as the first frame, last frame, starting frame, or ending frame. The routing contract says: if the images carry any reusable character, object, scene, or style responsibility, use the full-reference consistency skill instead. A minimal usage flow:

user: 'Use this as frame one and this as the final frame.'
agent: minimax-h3-creative-director checks for reference duties
       -> none found
       -> minimax-h3-keyframe-video-prompt builds boundary-frame prompt

Compare before adopting

Before replacing a generic video prompt strategy, compare this skill with minimax-h3-reference-video-prompt on the same boundary-frame request. Look at whether the generated video actually reaches the last frame. Check quality signals like issue activity, update freshness, and whether the author keeps the skill aligned with MiniMax H3 API changes. Also evaluate workflow fit: it is designed to sit downstream of the creative director, not to replace it.

Why it is worth tracking

This skill is part of a focused, actively maintained open-source toolkit for MiniMax H3 agents, with 105 stars and a narrow scope that is easy to validate. It is worth tracking because it solves a real routing ambiguity: many agents fail at literal boundary-frame requests simply because they do not distinguish an endpoint image from a reference image. For a centralized view of this skill and related MiniMax H3 skills, check its listing at https://www.openagentskill.com/skills/unknowlei-minimax-h3-keyframe-video-prompt. If you already use the creative-director skill, evaluate this one as soon as users start sending first/last frame image pairs.