MiniMax-H3 B-roll Generator

Turn a transcript or talking-head video into an approval-gated B-roll production plan and, when a source video is supplied, a timestamp-driven final edit. Preserve the visual grammar of Guangjun's T01–T21 templates and preserve the original spoken audio throughout the final assembly.

Non-negotiable gate

• Never submit a paid API request before the user has reviewed the shot list and explicitly approved specific shot IDs.
• Never interpret “continue”, “looks good”, silence, or approval of the workflow as approval to generate videos.
• Require all three conditions for a real API call:
  1. The current conversation contains explicit approval for named shot IDs or “all shots”.
  2. plan.approval.status and each selected shot's approval status are approved.
  3. scripts/minimax_h3.py is invoked with --approved.
• Never approve or submit a shot while a linked required asset is still marked missing.
• API keys must come only from MINIMAX_API_KEY. Never store a key in a plan file, prompt, review, manifest, or source code.
• Never let generated B-roll replace, regenerate, or time-stretch the source speech track.
• Every generated shot must forbid music, melody, singing, voice-over, and dialogue. Permit only synchronized foley and restrained environmental sound. If a render contains music, reject it or mute its audio before editing.

Required references

• Read prompt-versions.md at the start of every invocation.
• Read template-routing.md before classifying or styling shots.
• Read extended-template-library.md whenever T11–T21 may apply.
• After the user selects a profile, read exactly the matching complexity reference: template-complexity-v1-simple.md, template-complexity-v2.md, or template-complexity-v3-balanced.md.
• Read asset-intake.md before asking for, recording, or validating external assets.
• Read prompt-system.md before writing prompts.
• Read plan-schema.md before creating or editing a plan.
• Read minimax-api.md before any API operation or troubleshooting.
• Read video-input-and-editing.md whenever the user supplies a talking-head video or requests automatic editing.

Workflow

0. Choose the input mode

• Before asking for the prompt profile, ask whether the user wants to provide 逐字稿 or 口播视频, unless one is already supplied.
• For a transcript, preserve the existing workflow and never invent timestamps.
• For a talking-head video, inspect the file, then run scripts/transcribe_video.py to create a timestamped transcript. Preserve the exact path in source.video_path and the transcription artifact in source.transcript_path.
• Review obvious transcription uncertainty, product names, English terms, numbers, and named entities with the user before routing. Correcting the transcript resets downstream prompts and approvals.
• Use schema 1.6 for new video-input or automatic-edit projects. Read video-input-and-editing.md.

1. Ask for the prompt version

• Before segmentation, routing, asset requests, prompt writing, or plan creation, ask the user to choose v1 简洁清晰, v2 丰富动效, or v3 平衡清晰（推荐）.
• Skip the question only when the user already selected one in the same request.
• Do not silently default, infer a choice from an older task, or write all three versions unless explicitly asked for a comparison.
• Record the choice as top-level prompt_profile using v1_simple, v2_rich, or v3_balanced. Use schema 1.6 for new plans; schema 1.5 remains valid for legacy transcript-only plans.
• If the user changes versions after prompts exist, rewrite prompts and complexity plans, reset all approvals to pending, regenerate the review, and rerun dry-run.

2. Segment the transcript

• Split by semantic beat, not punctuation alone.
• Prefer 3–10 second beats. Keep a longer sentence intact when splitting would destroy meaning.
• Preserve the exact transcript text for every segment.
• If timestamps are absent, use ordered segment IDs and leave times null; do not invent precise timestamps.
• If timestamps came from the talking-head video, only merge adjacent transcript segments. Set the merged start to the first start and merged end to the last end; never estimate or rewrite timing by reading speed.

3. Decide whether a visual is needed

Assign every segment one route:

1. A_ROLL — personal judgment, emotional turn, caveat, direct address, or a line whose credibility depends on the speaker.
2. REAL_EVIDENCE — product UI, article, paper, source, chart, quotation, logo, or factual proof. Never fabricate these with AIGC.
3. EXISTING_MEDIA — a real product/person/place/process already present in the user's library.
4. HYPERFRAMES — deterministic typography/evidence/compositing that must be exact, editable, transparent, or tied to real UI.
5. MINIMAX_H3_PACKAGING — the beat is best communicated by a complete T01–T21-style text-led editorial motion-graphics shot, with exact approved on-screen words and supporting visual material.
6. MINIMAX_H3 — physical action, atmosphere, conceptual metaphor, environment, role scene, transition plate, or a clean moving hero subject where text is not the primary information carrier.

Do not force B-roll onto every sentence. A segment needs B-roll only when the visual adds information, clarifies a process, provides proof, establishes context, or creates a necessary pacing reset.

4. Route through T01–T21

• Choose one template_id or NONE for each non-A-roll segment.
• Respect the fallback rules in template-routing.md.
• First decide the primary information carrier: TEXT_PACKAGING, CINEMATIC_PLATE, REAL_EVIDENCE, or A_ROLL.
• Prefer MINIMAX_H3_PACKAGING for a hook, central concept, short conclusion, verified number, simple relation, chapter overview, or parallax statement whose meaning would be lost without visible words.
• Use MINIMAX_H3 only when action, place, object, atmosphere, or spatial behavior carries the meaning without typography.
• Keep brand marks, quotations, citations, product UI, long evidence, exact charts, and true-alpha overlays in Hyperframes or real capture.

5. Audit and request assets

• Inspect all user-supplied files and known library paths before asking for anything.
• For every non-A-roll segment, identify whether it needs a talking-head plate, product video/image, UI recording, logo, evidence document, reference image, portrait, map data, timeline data, diagram topology, technical specifications, font, brand guide, or background plate.
• Add every needed item to top-level asset_requests using plan-schema.md.
• Group missing requests into required blockers and optional fidelity improvements. Ask once, concisely, with the affected segment/shot, reason, and acceptable format.
• Never use MiniMax-H3 to fabricate a missing proof source, real product result, UI, logo, historical portrait, location claim, chronology, or technical relationship.
• If the user explicitly waives a required asset, change the concept to an obvious metaphor and require AIGC disclosure.

6. Write the generation prompt

• Write one continuous shot per generated clip; do not describe a montage.
• Apply the selected profile's phase, element, text, motion, camera, noise, and reading-hold budgets from prompt-versions.md.
• Describe subject, action, environment, composition, depth, camera path, lighting, materials, color, timing, and end state.
• Use at most three compatible camera instructions.
• For cinematic_plate, create clean negative space and explicitly forbid readable text.
• For motion_graphics or hybrid_packaging, declare the exact on_screen_text, hierarchy, position, font appearance, tracking, line height, colors, material layers, text entrance, profile-specific reading hold, camera curve, and end state. Explicitly forbid every other word and every glyph-like background texture.
• Keep H3 typography within the selected profile's text budget. Otherwise split the shot, generate a no-text plate, or route to Hyperframes.
• Never ask H3 to invent supporting copy. All visible wording must come from the transcript or user-approved structured data.
• When a supplied image is essential to composition or identity, prefer image_to_video and link its asset request; do not substitute a text-only approximation.
• When public reference video, audio, or image URLs are essential, use multimodal_to_video with reference_media; keep identity-critical real assets out of text-only approximation.
• Keep the final API prompt under 2000 characters.
• For MiniMax-H3 v2, default to 16:9, 2K, 6 seconds unless the user changes the production settings. V2 accepts only 768P or 2K.
• End every generated-shot prompt with an audio contract that allows only specific synchronized foley and restrained environmental sound. Explicitly forbid background music, melody, beat, singing, speech, dialogue, and voice-over.
• Add audio_design.music: false, audio_design.dialogue: false, a non-empty audio_design.sfx list, and audio_design.generated_audio: "sfx_only" to every shot.

7. Create the approval artifacts

Create both:

• broll-plan.json following plan-schema.md.
• broll-review.md generated by the review tool.

Validate and render the review:

python3 scripts/plan_tool.py validate broll-plan.json
python3 scripts/plan_tool.py review broll-plan.json --output broll-review.md

Show the user the recommended route, reason, prompt, mode, duration, resolution, and estimated number of paid generations. Stop and wait for approval.

8. Record explicit approval

Only after the user explicitly approves named shots, run:

python3 scripts/plan_tool.py approve broll-plan.json \
  --shots B001,B003 \
  --confirmation CONFIRM_MINIMAX_H3_COST

Use --shots all only when the user explicitly approves all generatable shots. Re-run review and show the approved subset before submission when approval wording is ambiguous.

9. Dry-run before spending

Always inspect the exact outgoing payloads first:

python3 scripts/minimax_h3.py broll-plan.json \
  --output-dir generated-broll \
  --dry-run

Dry-run does not require an API key and never accesses the network.

10. Generate approved videos

After the dry-run is correct and the approval remains valid:

python3 scripts/minimax_h3.py broll-plan.json \
  --output-dir generated-broll \
  --approved

The script submits sequentially, polls task status, downloads results, and writes generation-manifest.json. MiniMax-H3 uses v2; non-H3 legacy models retain the v1 path. Report failures without silently retrying a different model or changing prompts.

11. Inspect and assemble the final edit

• Visually inspect every generated shot for text, continuity, flicker, and unwanted music before assembly. A render containing music fails the sfx_only contract.
• For a talking-head project, run scripts/assemble_edit.py only after all selected shots have downloaded and passed review.
• Use the segment start_sec and end_sec values as the only edit timing source. Default to fullscreen_replace; keep A_ROLL segments untouched.
• Preserve the original talking-head audio at full level. Mix only approved B-roll foley/environment at the plan's low SFX gain. Never use generated dialogue or music.
• If a shot's audio is uncertain, pass --generated-audio mute; the final edit will keep the original speech without that shot's sound effects.
• Export the final MP4 and edit-manifest.json. Verify duration, audio presence, frame size, and that every inserted shot matches its approved time range.

Quality rules

• Enforce the selected profile mechanically through schema 1.6 for new plans; never mix v1 text limits with v2 motion density or label a v2 plan as v3.
• All profiles keep one primary semantic focus, exact text whitelisting, a stable reading end state, and deterministic fallback for brand/evidence-critical text.
• High-density templates should use a deliberately composed image-to-video first frame or supplied real media whenever identity, evidence, or layout matters.
• Favor editorial paper, dark archive, real material texture, controlled red annotation, and restrained blue accents.
• Keep the reading zone stable. Avoid constant rotation, elastic camera motion, and linear robotic pans.
• Use continuous eased camera movement and layered depth for T09-derived shots.
• Do not add decorative English, fake parameters, fake research, fake product screens, or meaningless microcopy.
• A text-led shot must contain at least one exact on-screen phrase and keep it sharp, near-front-facing, high contrast, and still enough to read for at least 1.2 seconds.
• MiniMax-H3 cannot guarantee exact local font files or error-free Chinese glyphs. Mark typography shots for text QA; reject and regenerate misspelled, duplicated, or invented words. Use Hyperframes when exact text is legally, factually, or brand-critical.
• Do not generate identifiable people without appropriate permission.
• Flag any shot that could be mistaken for documentary evidence as aigc_disclosure_required: true.

Failure handling

• If MiniMax-H3 is rejected by the endpoint, stop and show the exact API response. Do not automatically substitute another model.
• If H3 rejects a TokenPlan/Credit key, require a pay-as-you-go interface key with H3 access before retrying.
• Override only when the user supplies a valid replacement: MINIMAX_VIDEO_MODEL or plan-level api.model.
• If a task fails moderation, preserve the failed task record and ask the user before materially changing the concept.
• If the download URL expires, retrieve the file metadata again using the existing file_id; do not regenerate the video.