Sound Effects Planner for Shorts

suggest-sfx — the SFX pass

Take a rendered, voiced short and add tasteful sound effects, synced to the visual beats and the narration, drawn from a shared, growing SFX library. Work collaboratively and beat-by-beat; the user audits the plan and gives notes.

The shape: a declarative plan (sfx-plan.json) is the source of truth → a tool consumes it (tools/mix_sfx.py) → a library grows (media/library/sfx/) → a hard USER-AUDIT gate before anything is mixed.

Sound taste is a brand contract — read brand.md §7 every time. The house style is calm/premium (Linear / Anthropic / Vercel), felt-not-heard, always under the voice. NOT MrBeast-loud.

Inputs (read these first, every session)

• brand.md §7 (Sound design) — the SFX taste contract: subtlety, density, palette, levels, sync, signature motifs, source policy. Non-negotiable.
• shorts/short-N-<niche>/beats.json — the beats and REAL per-word voice times (written back by gen_voice.py). Sync cues to these words (a pop on "boom", a chime on "free").
• The shot source remotion/src/shots/short-N/*.tsx — the INTERNAL animation frames are where the real visual beats are (piece lands, digit pops, badge stamps). Read the shot, convert its local frame to global time: at_s = (local_frame + sequence_from) / fps. Do not guess — the cue must land on the exact frame the thing happens.
• media/library/sfx/catalog.json + palette.json — the library you draw from and grow.

The library (media/library/sfx/) — the durable asset

• palette.json — generation recipes: generic, reusable sound ids + prompts + duration + tags. The source of truth for what the library SHOULD contain.
• catalog.json — the manifest of what EXISTS: {id, file, category, tags, duration_s, peak_dbfs, loudness_lufs, source, model, license, prompt, used_in} per clip. Written by gen_sfx.py.
• clips/*.mp3 — loudness-normalized clips (~−20 LUFS, −1.5 dBFS ceiling) so a plan's per-cue gain_db is perceptually meaningful.
• Library-first, always. Reuse an existing clip before generating. Name/tag clips GENERICALLY (ui-toggle-on, whoosh-soft, pop-reveal) so future videos reuse them — the library is the durable asset, each video is one draw from it. Only genuinely-missing sounds get added to palette.json and generated.

Workflow

1. Read brand §7 + beats.json + the shot source + catalog. Pull exact cue times: for each candidate beat, find the visual moment's local frame in the shot and/or the narration word, and convert to global seconds.
2. Decide taste with the user up front if unsettled (present-ness, density, which kinds, source) — then apply brand §7. Don't re-ask settled decisions.
3. Propose the cue list — function-first. For each beat ask what it needs (brand §7): motion (whoosh) · tension (riser) · emphasis (impact/pop) · snap (click). Score key transitions, reveals, and tasteful click-sequences; a 3–4-shot click-sequence counts as ONE gesture. Mark genuinely deniable texture as "optional": true. Layer the 2–3 biggest moments (build-and-drop): riser→impact on a scripted reveal, whoosh→pop so a cut stands out — in the plan a layer is two events at the same/adjacent at_s that sum. Sync each cue to the VISUAL beat and often the exact word.
4. Library-first sourcing. For each distinct sound the plan needs, reuse a catalog clip if one fits. For misses, add a generic recipe to palette.json and generate: python tools/gen_sfx.py (ElevenLabs Sound Effects API → normalized clip → catalog). Fall back to curated royalty-free only where generation is weak; record source + license either way.
5. Author shorts/short-N-<niche>/sfx-plan.json — one event per cue, at_s in global seconds. (Schema below.)
6. USER-AUDIT GATE (hard). Present the resolved cue sheet (python tools/mix_sfx.py <plan> --print) and get the user's approval/notes BEFORE mixing.
7. Mix the audition preview once approved: python tools/mix_sfx.py shorts/short-N-<niche>/sfx-plan.json → the plan's render.out. Light sidechain duck under the voice + a safety limiter. Iterate on the user's notes (gains, timing, add/cut cues) — re-print, re-mix.
   • Verify audibility with numbers, not hope. After mixing, RMS-diff the mixed audio vs the voice-only preview at each cue window: a story-critical cue should add ≥ +4 dB, texture +1–3 dB. Inaudible cues hide in cue sheets — don't ship a cue you haven't confirmed lands.
   • Transient-clip gain gotcha. Short percussive clips (knock/stamp/keys/snap) hit the −1.5 dBFS peak ceiling BEFORE reaching the −20 LUFS loudness target, so they catalog ~3–5 dB quieter than sustained clips. Their plan gain_db must be ~3–5 dB higher than the brand table implies. The audibility check above is what surfaces this.
   • SHORTS calibration. Under a short's near-continuous narration (~2.7 w/s, few gaps) a conservative gain table is 4–8 dB too quiet across the board — the duck + voice masking eat everything. Proven landing zone: transitions/whooshes −3, story pops/impacts 0..+3, stamps/snaps 0..+7, layered-hero risers 0..+2. Start a short's plan there.
   • Cues that sit fully UNDER continuous speech never measure — accept felt-not-heard or cut them. The sidechain duck suppresses quiet clips the whole time the voice is active (a click-sequence can measure +0 dB at ANY gain). Set such cues at a sane +6-ish and let them live in the word gaps, or delete them; do NOT chase them with gain (a pause would make them spike).
   • Measure transients with tight windows. RMS over 0.6s dilutes a 50 ms click to nothing and an adjacent loud cue can leak in and fake a pass. Use ~0.3s windows for snap/pop/zap/stamp, and measure a riser at its final third (its energy is at the END).
   • A noisy TEXTURE cannot be fixed by gain. Static/glitchy/hummy clips read as NOISE over speech even at −4 dB; lowering them just makes quiet noise. If the user says "noisy", swap the sound's CHARACTER (clean mechanical snap) or use silence — see brand §7 "no static/glitch textures under narration".
8. Save back. New clips stay in the library + catalog (they carry to the next video). Update used_in.