| name | sound-cue-designer |
| description | Use when planning sound effects, audio cues, sonic logos, notification-like prompts, background music, or mix boundaries for AI paper explainer videos before TTS merge, captions, HyperFrames render, or platform review. |
Sound Cue Designer
Core Rule
Use sound cues as auditory bookmarks for comprehension, not as decoration. A cue is valid only when it marks a real cognitive turn such as hook, formula reveal, Q/K/V card entry, softmax normalization, weighted V aggregation, engineering-layer shift, or next-episode CTA.
Required Checks
- Do not overpower voiceover. Keep short effects about
12-18 dB below spoken voice; keep background music about 18-24 dB below spoken voice.
- Avoid phone notification sounds, alarms, game sounds, meme stingers, harsh beeps, or dense template sound packs.
- Keep high-frequency cues away from English terms such as
Attention, softmax, FlashAttention, GQA, MQA, KV Cache, and vLLM.
- Place cues between phrases or on visual action points; do not mask consonants, numbers, formulas, or subtitles.
- Treat the final mix as part of the TTS gate: if effects hurt ASR transcript diff, English clarity, or human listening review, lower or remove them.
- Do not add cue instructions into
spoken_text; keep them in storyboard, FRAME, sound cue plan, or HyperFrames prompt.
- Require a final mix status when the episode contract includes SFX. A video build must not silently fall back to a no-SFX voiceover after sound cues are approved.
- Add special listening checks around professional terms, formulas, and next-episode CTA when a cue lands near those phrases.
Paper Explainer Cue Map
| Moment | Cue | Purpose |
|---|
| Opening | restrained sonic logo | Establish series identity |
| Core formula | QK reveal | Mark the episode's main operation |
| Concept cards | Q/K/V card taps | Help memory without adding narration |
| Scaling | compression or squeeze cue | Indicate sqrt(d_k) stabilization |
| Softmax | softmax normalization rise | Convert scores into distribution |
| Value read | weighted V aggregation merge | Show many streams becoming one output |
| Engineering layer | three subdued card hits | Separate compute, architecture, runtime |
| CTA | gentle upward tail | Lead into the next episode |
Output Contract
When used for an episode, produce or update an episode-level sound cue plan with:
- cue id and spoken cue it supports
- visual action it syncs to
- sound type, duration, and relative loudness
- risk to voice clarity, ASR transcript diff, captions, or platform tone
- verification note for human listening review
- final mix expectation: required/optional/off, target mixed file, and fallback policy
Boundaries
- Do not synthesize, download, or license sound assets.
- Do not mix final audio or replace voiceover.
- Do not change script claims, captions, formulas, or spoken text.
- Do not use sound cues to compensate for unclear narration; fix the script first.
- Do not approve a final MP4 as SFX-complete without evidence that the mixed audio file, not the dry voiceover file, was used.