comfy-agent minimax-h3-prompting — MiniMax H3 Skill
为 comfy-agent 的 H3 预设写和迭代 Prompt，专门处理音频不对、运镜被忽略、风格漂移和节奏过快。
为 comfy-agent 的 H3 预设写和迭代 Prompt，专门处理音频不对、运镜被忽略、风格漂移和节奏过快。
npx skills add https://github.com/shinshin86/comfy-agent/tree/main/.claude/skills/minimax-h3-prompting
shinshin86/comfy-agent · MIT · ★ 14 (2026-09-13)
原文: comfy-agent minimax-h3-prompting
shinshin86 · 出处
Write or refine prompts for MiniMax H3 video generation via comfy-agent (minimax_h3_t2v / minimax_h3_i2v / minimax_h3_r2v presets). Use whenever composing the --104_prompt for an H3 run, when a user asks for a video clip on an H3 environment, or when an H3 output missed the intent (wrong audio, ignored camera move, style drift, rushed pacing) and the prompt needs iteration.
minimax-h3-prompting
Read docs/minimax-h3-prompting.md
before composing any H3 prompt — it is the single source of truth.
Non-negotiables it enforces:
- H3 renders picture and stereo audio in one pass: always direct
ambience and music explicitly, or the model invents them.
- Render the final block in the official three-field format
(integrated_multimodal_description / overall_soundscape /
non_diegetic_music), with the <Picture 1> alignment sentence for I2V or
explicit <Picture 1> / <Audio 1> role assignment for R2V.
- Fit beats to the frame budget (--104_length: 124 ≈ 5 s → one beat;
243 ≈ 10 s; 362 ≈ 15 s → timecoded shot list).
- Iterate with a fixed --15_noise_seed; verify frames (ffmpeg) AND
audio (ffprobe) before reporting results.