intro-video

Intro Video

Build Remotion intros, reels, and brand films with remocn. Load the remocn skill for components. This skill owns captions, audio, intros, and render.

Three rules, each learned by shipping the wrong video:

1. Caption timing comes from ASR. Never estimate from character counts or energy-based pause detection. Estimating ran 1.1–1.9s late on a ~20s reel. That is a wrong video.
2. Trim the silence off both ends of the VO before ASR. See Audio.
3. Overlays live in the empty top-left, not across the frame, unless the user asks otherwise. See Overlays.

Inputs

Ask for both before any caption work:

• Script — spelling. ASR mishears proper nouns ("GROG", "MIND CRAFT").
• Voice audio (or a video that already has the VO).

Script gives spelling. ASR gives timing. Map script words onto ASR timestamps; keep the times.

Audio

Trim the dead air off both ends before anything else. Raw VO arrives with silence at the head and tail. Ship it untrimmed and the video opens on a still frame with no voice and ends hanging on nothing.

Order matters: trim first, then ASR. ASR timestamps are relative to the file you feed it. Trim afterwards and every caption is offset by the length of the head silence.

Measure the real boundaries — do not guess a fixed number:

ffmpeg -i voice.wav -af silencedetect=noise=-40dB:d=0.15 -f null - 2>&1 \
  | grep silence_

Read the first silence_end (voice starts) and the last silence_start (voice ends), then cut to those with a small margin so consonants survive:

Confirm with ffprobe -show_entries format=duration before moving on. Then run ASR on voice-trim.wav, and use that same file in the composition — the timings only line up against the file they were measured from.

• -40dB suits a clean close-mic recording. Noisy room, raise toward -30dB.
• Leave ~0.05s ahead of the first word and ~0.10s after the last. A hard zero cut clips plosives and sounds truncated.
• Normalise to −16 LUFS after trimming.
• ~0.2s crossfade at clip joins. Build the whole track in one ffmpeg filtergraph so frame counts match exactly.

ASR

Uses sherpa-onnx from PyPI and a zipformer model from GitHub release assets (allowlisted). Whisper weights on Azure/HuggingFace are often blocked — do not start there. The script downloads the model on first run.

Drive every caption from / durationInFrames from that JSON. 2–4 words on screen. Keyword in amber.

Craft

• Opening clip: vignette + Ken Burns 1.0 → 1.09, origin 50% 42%.
• Logos: real assets only — npm icon packs or the product's own GitHub repo. Never redraw.

Overlays

Animations, logo lockups, code cards, stat callouts and every other overlay go in the empty space in the top-left. Never across the whole screen — unless the user explicitly asks for full-screen.

A full-bleed animation covers the speaker and fights the captions. The top-left is empty on nearly every talking-head frame, so that is where an overlay reads without hiding anything.

Look at an actual frame first and put the overlay in the empty corner you see there. Where the subject genuinely sits left, mirror the box to the top-right. Starting points:

• Keep the overlay clear of the caption band at the bottom and, on vertical, of the right-hand action rail.
• Animate in place — scale/fade/slide within the box. Do not let a transition sweep across the full frame on the way in.
• One overlay at a time. Two competing corners is worse than none.
• Full-screen is legitimate for a deliberate cutaway or end card. That is a different beat, not an overlay — cut to it, do not lay it over the speaker.

Render

• Pin typescript@5. TS7 drops ts.sys and breaks Remotion's bundler.
• @/ needs a webpack alias in remotion.config.ts.
• If Chrome Headless Shell cannot download, extract Chromium from @sparticuz/chromium (brotli).
• Render 150–225 frame chunks, concat, then mux audio. Full renders blow tool timeouts.
• Concurrency 1 on a single core.
• If remocn.dev 403s, pull components from the remocn GitHub repo registry-artifacts/ instead of shadcn add. npm, PyPI, and GitHub release assets work.
• Frame format JPEG, quality 95. PNG at 4K is ~30MB/frame — a 200-frame chunk needs ~6GB of temp space and fails silently when the disk runs out. Check df -h / and clear out/ first.
• A chunk that overruns leaves a truncated file with no error. ffprobe the frame count after every chunk before concatenating.

4K

Render the 1920×1080 composition with --scale=2. It stays 1920×1080 logically but rasterizes at 2× device pixels, so captions, lockups and vignettes come out natively sharp. Do not build a separate 4K composition.

Concat the chunks, then mux the audio.

• Re-cut every source asset at native 4K first.
• ~74 frames per chunk over talking-head footage. Drop to ~40 over cutaways that are themselves 4K video — compositing 4K over a decoding 4K clip is much slower.
• 1080p plus an ffmpeg lanczos upscale takes ~4 min against ~1 hour for true 4K. Offer it, but confirm before switching.

Cover art

Safe areas differ per platform and the wrong guess buries the type. See references/cover-art.md.