Minimax H3 Keyframe Video Prompt

MiniMax H3 Keyframe Video Prompt

Design a continuous visual path from or toward concrete boundary frames. Treat a boundary image as an actual target frame, not merely a loose style reference.

Routing Contract

New MiniMax H3 requests should enter through minimax-h3-creative-director. Use this skill only for explicitly declared boundary-only images. If the user did not explicitly say first frame, last frame, starting frame, ending frame, or an equivalent literal boundary role, use minimax-h3-reference-video-prompt. If any image also defines or preserves a character, identity, person, object, costume, scene, style, voice, action, camera, composition, or other reusable trait, use full-reference even if the user also calls it a first or last frame.

Official Format Authority

Before drafting, read ../h3-prompt-writing/SKILL.md and ../h3-prompt-writing/references/base-en.txt. Treat those official MiniMax files as the canonical prompt-format specification. If this skill or its local references conflict with them, follow the official files.

Read references/keyframe-rules.md before drafting.

Confirmed Multishot Handoff

When the director supplies a confirmed multishot_plan, treat its shot count, timing, content, framing, performance, camera, transitions, sound, and continuity decisions as already answered. Do not ask those questions again. Map every confirmed shot into the official keyframe prompt while preserving exact boundary-frame alignment. Reopen a decision only when the plan is physically impossible, contradicts a boundary image, or exceeds the effective duration.

Mandatory Format

Always return the exact mode-specific picture-alignment instruction followed by the official three fields. Never return a free-form natural-language paragraph, keyword list, shortened prompt, or alternate schema as the final English prompt.

Interactive Direction Check

Before drafting, use the host's structured-choice tool whenever the eligible pure-keyframe brief is sparse, the intended transformation is underspecified, two or more of duration, motion path, visual treatment, camera strategy, dialogue/sound, or ending behavior are missing, or the user delegates creative direction to the AI.

In OpenCode, call the built-in question tool rather than printing Markdown options. In Codex, use the available structured user-input tool; if unavailable, ask one concise plain-text question.

Ask as many related questions as the current decision stage requires; normally 1-5, but three is not a per-call, per-turn, or per-session ceiling. A strict yes/no question may contain exactly two options; every other choice question must contain at least five materially different, feasible options. Put the context-specific recommendation first with (Recommended) in its label, explain each consequence in one sentence, omit Other, and enable multiple selection only for genuinely combinable choices. Before every question tool call, count the options in the actual payload. If any non-binary choice has fewer than five, do not submit it; expand it with meaningful alternatives or make it open-ended. Do not create near-duplicates merely to reach five.

Prioritize:

• Boundary role when one image is ambiguous: first frame or last frame
• Motion strength: subtle natural motion, clear narrative action, dramatic transformation
• Camera strategy: continuous single shot, restrained cuts, dynamic montage
• Landing behavior: exact stable hold, arrive on the final instant, transition through the frame

For a sparse or delegated brief, ask enough high-impact questions in one batch to establish the current stage, normally 3-5. Do not treat "improvise" as permission to skip questions. Skip only when the user explicitly prohibits questions, then disclose the chosen direction.

Select the Mode

• Explicit first-frame-only image with no reusable reference role -> I2VA.
• Explicit first-and-last-frame pair with no reusable reference role -> FL2VA.
• Explicit last-frame-only image with no reusable reference role -> L2VA.

If boundary wording is absent, ambiguous, or merely inferred, do not ask the user to confirm this specialist. Route directly to minimax-h3-reference-video-prompt.

If the same boundary image or any other asset must also control identity, character, object, action, style, scene, sound, editing, or pacing, use minimax-h3-reference-video-prompt and express the first/last-frame relationship inside the full-reference prompt. Do not mix API first_frame/last_frame roles with reference_* roles.

Workflow

1. Confirm the boundary role, exact integer duration, and intended final action or transformation. Keep the official online target within 4-15 seconds, 24 FPS, and 7000 prompt characters; obey narrower local-workflow limits when known.
2. Inspect supplied images when available. Record identity, clothing, pose, object states, composition, camera angle, lighting, and spatial relationships.
3. Define the motion path before writing prose.
4. Keep subject identity and persistent attributes consistent unless the user explicitly requests transformation.
5. Prefer a single continuous transition for FL2VA and usually for L2VA. Two boundary images do not automatically create a cut. Add cuts only when explicitly requested or essential.
6. Write the exact alignment instruction, then the three core fields.
7. Validate that the target boundary frame is reached at the exact endpoint rather than early.
8. Return the required bilingual output.