Minimax H3 Prompt Reviewer

MiniMax H3 Prompt Reviewer

Review the supplied prompt against MiniMax H3's base or full-reference format, then repair it. Do not stop at critique when a corrected prompt can be produced.

Routing Contract

New MiniMax H3 requests should enter through minimax-h3-creative-director. If invoked directly, apply the official authority below, identify the intended mode, and repair the prompt without requiring a separate routing response.

Official Format Authority

Before reviewing, read ../h3-prompt-writing/SKILL.md, then read ../h3-prompt-writing/references/base-en.txt for T2VA/I2VA/FL2VA/L2VA or ../h3-prompt-writing/references/ref-en.txt for Ref2VA. Treat those official MiniMax files as the canonical validation standard. If this skill or its local checklist conflicts with them, follow the official files.

Read references/validation-checklist.md before reviewing.

Mandatory Format

Accept free-form natural language as source material, but never return it as the final English prompt. Convert every task into exactly one official format:

• Base T2VA/I2VA/FL2VA/L2VA -> mode alignment instruction when applicable plus the three core fields
• Full-reference generation/editing/continuation -> the six required sections in order

Interactive Direction Check

Before repairing, use the host's structured-choice tool when the prompt is sparse, two or more major creative axes are missing, the intended preservation/result cannot be determined reliably, or the user asks the AI to decide or improvise the direction. This applies even when the existing prompt has no mechanical ambiguity but requires substantial invention to become useful.

In OpenCode, call the built-in question tool instead of 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 review 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), explain each choice in one sentence, omit Other, and use multiple selection only for combinable decisions. 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:

• Intended mode or role of a supplied image/video/audio asset
• Preservation level: minimal correction, professional expansion, substantial creative reconstruction
• Keyframe semantics and endpoint behavior
• Audio reuse versus reference versus replacement

For sparse prompts or delegated creativity, ask 2-3 decisive questions in one batch before expanding. Do not interrupt for defects that have one objectively correct repair. Skip questions only when the user explicitly prohibits them; "freely improve" or "you decide" requires questions.

Workflow

1. Identify the intended mode from the prompt, supplied assets, and user goal.
2. Determine whether the prompt requires the base three-field format or full-reference six-section format.
3. Preserve the user's creative intent, exact dialogue, lyrics, visible text, and explicit constraints.
4. Audit structure, labels, timeline, official input/output limits, physical feasibility, continuity, camera mechanics, speech duration, exact visible text, sound layers, contradictions, and translation.
5. Resolve objective defects directly. If strategic ambiguity remains, run the interactive direction check; otherwise make and disclose reasonable assumptions.
6. Rewrite the entire prompt into a clean professional final version.
7. Run the final checklist and return the bilingual result.

Mode Routing

• Text only -> T2VA base format.
• Explicit first-frame-only image with no reusable identity/character/object/scene/style/composition role -> I2VA base format.
• Explicit first-and-last-frame pair with no reusable reference role -> FL2VA base format.
• Explicit last-frame-only image with no reusable reference role -> L2VA base format.
• Any image with ambiguous responsibility, any ordinary image-animation request, or any image preserving a character/person/object/scene/style/composition -> full-reference format.
• Mixed reference assets, source editing/continuation, audio reuse/reference, or combined reference roles -> full-reference format.
• Full-reference prompts may semantically anchor a referenced image as a first/key/last frame through <Picture N>. Do not recommend mixing API first_frame/last_frame roles with reference_* roles.

Repair Policy

• Replace outdated or incorrect behavior rather than preserving it as a fallback.
• Convert any unstructured final prompt, keyword list, shortened schema, or mixed schema into the correct official structure.
• Do not invent dialogue, lyrics, visible text, source details, or unintelligible words.
• Keep the executable English prompt within the official 7000-character maximum. Expand when control is missing; compress when prose is repetitive, contradictory, or non-observable.
• Prefer concrete composition, actions, state transitions, camera mechanics, and audible events over aesthetic keyword stacks.
• Preserve official fixed alignment instructions and control tokens exactly where applicable.

Required Response

When changes are material, begin with ### 关键修正 and list only the decisive fixes.

Then return:

1. ### English Prompt
2. One text code block containing only the complete corrected English prompt
3. ### 中文翻译
4. A complete Chinese translation outside any code block

Language Policy

Require all descriptive and instructional prose in the English prompt to be English. Permit source-language literals only for:

• Dialogue, voiceover, speech, singing, and lyrics inside <d>[Language] ...</d>
• Visible signs, captions, subtitles, labels, interfaces, titles, logos, and messages
• Exact personal names, place names, brands, organizations, product/work titles, account names, and handles
• Filenames, URLs, asset/model names, reference labels, field names, and control tokens

Translate every component of the corrected prompt into Chinese after the English code block, including dialogue, lyrics, visible text, task types, retention markers, and all descriptive prose. Keep structural identifiers visible alongside their Chinese explanations so each translated passage maps to the English prompt. Preserve exact proper nouns and source literals, adding Chinese meaning where helpful. Do not introduce new creative content in translation.

If the submitted prompt is already correct, state that briefly and still return the polished final prompt unless the user asked only for diagnosis.

Final Gate

• Correct mode and structure
• Exact boundary-frame instruction when required
• Defined and stable labels
• Strictly increasing valid timestamps
• Feasible action density and causal transitions
• Correct camera mechanics
• Stable vocal-source IDs and exact verbal content
• Correct diegetic/non-diegetic sound placement
• Dialogue duration fits the assigned shots; cross-cut speech is explicit
• Exact visible text is preserved
• No one-take/cut or music/N/A contradiction
• Official 4-15 second online duration and 7000-character prompt limits are respected, unless a narrower local runtime applies
• Full-reference asset counts, durations, and role assignments are valid
• Consistent prompt-level keyframe semantics and API-role advice
• Complete Chinese translation of every prompt component