gordonlu/awesome-minimax-h3-skills

H3 Promo Film Studio

Use this Skill when the user wants a promo or trailer from MiniMax H3 from text alone: "make a 6-second Peking-duck promo", "a product launch film", "a sci-fi teaser for my channel". No images, no clips — the prompt is the entire production.

Core principle: the description is the footage. Every visible beat must exist in text with a timecode, or it will not exist in the video.

Prompts follow h3-prompt-writing (base-en.txt). Field names, order, and timing notation come from the guide and must not be improvised.

1. Trigger

Use when: pure-text generation requests in any language; any of T2VA's natural jobs — food / product promos, trailers, ambient or cinematic shots, explainer-style one-shots, kinetic-type pieces.

Do not use when: the user supplies reference images (I2VA/FL2VA), reference videos (Ref2VA), or wants to edit/continue an existing clip. Route to h3-keyframe-film or reference-motion-transfer instead. Requests for >15s continuous output require segmentation.

2. Intake

Confirm: ① concept & mood (one line), ② aspect ratio (16:9 / 9:16 / 1:1 / 4:3 / 3:4 / 21:9), ③ duration (4–15s; default 6s), ④ visual style (cinematic / hand-drawn / ad / 2D…), ⑤ on-screen text (exact wording & language), ⑥ audio policy (silent / SFX / BGM direction).

3. Goal

One approved video whose every beat matches the written timecoded description — scene, camera, subject, and audio all visibly derived from the prompt, with no invented content.

4. Analysis

Decompose the concept into 3–5 beats, one readable action per beat (≥1s each). Note the camera move per beat (push-in / cut / drift…), and flag anything the model cannot render from text (exact logos, real people, brand likeness) — decline those or abstract them.

5. Planning

• Beat sheet — timecoded, e.g. 0–2s wide vista → 2–5s hull close-up → 5–8s cockpit drift. Beats must tile the full duration; the last beat ends exactly at the total duration.
• Structure — one shot per beat transition or a continuous single shot; [Shot N] numbering in order.
• Style sentence first — the very first sentence of integrated_multimodal_description fixes style, lighting, and quality grade.
• On-screen text — state position in the frame and keep the original language.

6. Prompt Construction

Three fields, in this order:

1. integrated_multimodal_description — style sentence → [Shot N] headers (no timestamp) → timecoded beat prose: From X.XX to Y.YY seconds, .... Each beat: subject, action, camera, and any visible text.
2. overall_soundscape — diegetic sound per beat.
3. non_diegetic_music — genre, BPM, energy curve; N/A when unused.

Rules: no reference labels (<Picture N> / <Video N>) — this is T2VA; 350–500 English words for the description; dialogue/lyrics/on-screen text stay in their original language.

Ready-to-run examples: references/prompt-library.md.

7. Generation & Review

• Model: MiniMax H3, ratio & duration as planned. First pass default resolution; upgrade to 2K only after approval.
• Review checklist: beat coverage (every timecoded beat appears in order) / scene consistency / subject consistency / camera follows the written move / on-screen text correct / audio matches policy. Score pass / partial (name the beat) / fail.

8. Retry Strategy

Fix the cheapest layer first:

Symptom	Fix
A beat is missing or merged	Re-write that beat's timecode window explicitly; shorten duration
Camera doesn't match	State the move inside the beat sentence ("the camera pushes in")
Style drift	Re-state the style sentence verbatim as the first sentence
On-screen text wrong	Shorten the phrase; state exact position (e.g. "bottom right")
Generic/boring result	Add one concrete object detail per beat; tighten tempo contrast

Escalate after two failed retries of the same layer — offer a shorter window or fewer beats instead of a third blind retry.

9. Output

Deliver: final video file; the beat sheet; the exact prompt text used (reusable asset); settings record (ratio, duration, audio policy, resolution pass); review notes.

Boundaries

Text-only generation only. If the user mentions uploading an image or clip, stop and route to h3-keyframe-film / reference-motion-transfer. Do not generate real identifiable persons, brand logos, or copyrighted characters; abstract them or decline.