Short edit in Beacon

Approved on the Grok plugins short in ~/Movies/Beacon/3-shorts-new.beproj (2026-10-28; the user: "Done an amazing job"). Earlier hand-made shorts in ~/Movies/Beacon/3-shorts.beproj set the layout. Open either to copy from.

Order: plan → rough cut → camera → top half → highlights → logos → zooms → captions → audio → verify. Titles and descriptions afterwards: short-edit §6.

0. Before anything

• project_open with the project's path: the app often has two projects open and refuses a bare open.
• A project can hold several shorts (Grok, then Midterms). Edit only the one asked for. Beacon's Remove Retakes / Silences / Fillers act on the whole timeline: dry-run them, keep only the ranges inside the short, apply those (transcript_removeSilences with at, timeline_rippleDelete with ranges). Ripple edits move the later shorts left: fine, but their clips and the gap must stay as they were. Check afterwards.
• For anything beyond the rough cut, write the plan first and wait for the go (user request): a table of time → what fills the top half, the highlights, the B-roll count, caption placement, and the open questions.

1. Rough cut (shorts are tighter than long videos)

• Read transcript_get for the short's range. Keep the last full take of every line; cut false starts, off-mic chatter, and Whisper's hallucinated "Thank you" / "Okay" in quiet stretches. A short's script is small, so cut by hand with transcript_cut word ranges in one call, rather than trusting Remove Retakes' picks.
• Lists said in several takes ("Shopify / Benzinga / particle, Oriane and much more / Apollo, FullEnrich and much more"): keep the last ending and say which one was dropped.
• Silences: over 0.25 s, leave 0.2 s (measured on the user's shorts: ~0.1 s of quiet each side of a cut; long videos leave 0.3). Then trim the short's tail to ~0.1 s after the last word. Recompute times after every ripple: a range computed before a cut lands in the wrong place after it.
• Stretched words hide repeats and pauses (Whisper writes a word said twice once, and swallows pauses into words): dump transcript_get words, flag any word over ~2x its expected length, check its loudness from the .brec's microphone.wav (Beacon's ffmpeg has no silencedetect: raw PCM + Python RMS), and ripple-delete what's left. To hear what a stretched word hides, cut that stretch out with Beacon's ffmpeg and run Beacon's whisper-cli -ml 1 -sow on it alone (model in ~/Library/Caches/Beacon/models); captions over it need re-splitting and caption_setText by hand.
• A screen recording with voice-over (no camera) plays full screen; zoom the screen clip onto each item as it is named (scale and move about the point: pos' = P + k(pos - P)).
• A typed prompt (ChatGPT etc.) is a hand-written full-screen motion_code graphic of the app's home page, the text typed per word as it's said, input box in the middle of the screen, no captions over it.

2. Camera

Canvas 1080x1920. Transform coordinates: origin at the centre, y up, so the top half is y +480 and the bottom half y -480.

• Full-screen camera for the hook line and the CTA (the user's own edit, as in earlier shorts), and for one emotional beat in the middle ("communicate with people on my behalf"). Nothing goes over the top half then: no title graphic, no logo.
• Alternate bottom-half and punch-in across cuts: it hides the jump cuts. Keep the cap clear of the panel line (y -545, not -512).
• A slow push-in (0.45 → 0.5, x -44, y -500, easeInOut) on any take over ~8 s.
• Set the framing on one clip, frame_at, then apply it to the rest in one clip_set.
• Intro and outro moves (user request): a slow zoom in across the whole full-screen hook (scale 0.92 → ~1.06, linear, split across its clips so the speed carries over a cut) and a slow zoom out over the CTA (start zoomed, easeOut back to 0.92). Zoom about the face: pos' = P + k(pos - P).
• A small pop-in prop on a line about the speaker ("I'm not an American" got a waving US flag beside the cap): hand-drawn motion_code canvas, about 1 s, kept clear of the face and the frame edge at the backOut overshoot.

3. The top half (y +480, 1080x960): never empty while the camera is halved

Every second of a half-camera shot has something above it. After placing, list the top-half clips and check for gaps: a 3-frame gap showed black.

Screenshots (real pages, never the user's inbox or workspace):

• Browser width 400 CSS px (user rule; 300 for narrow pages) at 2x pixel density, iPhone user agent, so pages show their mobile layout: ~/skills/short-edit-beacon/scripts/screenshot.sh <out.png> <url> [height]. Headless Chromium from Playwright's cache; Agent Browser or Chrome over CDP also work.
• Crop to 800x711 (the panel's shape) from where the page's message starts, above any cookie banner: scripts/crop_panel.py <in> <top_px> <out> (save into the project's media/screens/). Place with media_import (at, a Screenshots track), clip_trim to the line, transform x 0, y 480, scale 1.35, and a slow push-in to 1.42.
• News sites dim the page under a cookie wall, and some (USA Today) block headless Chromium. Capture with the page's banners and full-screen fixed layers removed before the shot (scripts/shot-clean.ts), or pick another outlet. ChatGPT's plugin pages render blank without a login; use the vendor's own announcement page.
• Check a reused screenshot is what it claims (a "Firecrawl plugin in Grok Bot" shot was ChatGPT's apps page).

Highlighter wipes on the words being said (user favourite):

• edit_addAnnotation preset highlight, then clip_set region width 2, height 54 (body text) or 86 (headlines), and animate region.width and transform.x together (easeOut, 0.35-0.8 s, ending as the word ends) so the left edge stays put.
• From a 540-wide frame_at: canvas px = frame px x 2, then x = left + width/2 - 540, y = 960 - centre_y. Start x = left + 1 - 540.
• One bar per line for wrapped headlines, staggered 0.3-0.5 s.

Motion graphics: motion_add (time, duration, a typed prompt), then clip_set motionBox {1080, 960} and transform y 480 on all of them, then motion_make each (OpenAI, look Bold; ~2 min for 8 in parallel). Prompts name the exact words, numbers and brand colours (Firecrawl #FA5D19) and say "as it is said". What worked: a comparison bar chart, a counter racing to the number, names popping in one per word, an email drafted then "Sent ✓", channels collapsing into a Summary card, chat bubbles labelled per word.

B-roll for lines about a feeling or a real-world moment (paywall, "saved me a lot of time"): broll_add region [0, 0, 1080, 960] with a written prompt, broll_estimate, then broll_generate (Seedance 2.5, ~$0.22/s; the user allows as many as are best). Flag generic stock-person shots.

Full-page beats (user favourite, 2-4 s, hide the camera): a full-screen screenshot on the product's name (the vendor's home page, 400 CSS px x 711 tall, scale 1.35 at y 0), a logo wall on a list of categories ("travel to sports to finance and much more": a row of real provider app icons per word, then a zoom out into a wall and "+ much more"), a data graphic on a hook about a topic (a US map coloured by party with seat bars and counts under it). No captions over them; remove the caption clips in the span. Real data only: check the numbers that week and footnote what they mean. For Firecrawl, providers come from the Alexandria catalogue (firecrawl_find_tools, categories, level providers).

Beacon's motion graphics load nothing from outside the page, and a page with embedded icons or map paths is too big for one motion_code call. Render those yourself: write the page with window.__frame(t), then bun scripts/render-html.ts <page.html> <frames dir> <seconds> (headless Chromium, 1080x1920, 30 fps) and encode with Beacon's ffmpeg (-c:v hevc_videotoolbox -q:v 80 -tag:v hvc1 -pix_fmt yuv420p) into the project's media/, then media_import onto the Motion graphics track. Check a contact sheet of the frames first. To replace one, import a new file name (Beacon keeps the old asset).

Logos: real ones only. App icons: the iTunes Search API (https://itunes.apple.com/search?term=<app>&entity=software, swap 512x512bb.jpg for 1024x1024bb.png in artworkUrl512); check the seller. On a Logos track at the top, round-corners + drop-shadow, popping in on the spoken name: scale 0.01 and rotation -12, then backOut to scale ~0.16 over 0.35 s. Sparingly: the user kept one (on the Gmail screenshot) and removed the others (hook, Slack).

4. Captions

captions_create with from/to = the short's range, maxWords 3, style: Montserrat 800, size 72, uppercase, white, highlight [1, 0.84, 0.1, 1], background [0, 0, 0, 0.75], radius 18, maxWidth 0.9, anchorY middle, offset 0 (on the seam between the halves; user choice). Captions over a full-screen camera shot go lower: caption_place each one (one clip per call; clips is refused) to middle, offsetY 101, and give them a blue label (selection_set, then edit_label). The track's highlight stays off (alpha 0): captions_create's style changes the whole track, so never pass a visible highlight. Leave a full-screen motion graphic with no captions: create captions on either side of it. Then fix Whisper's spellings with caption_setText (GrokBot, paywall, FullEnrich), and after any re-cut look for one-frame sliver captions.

5. Audio (last, once the picture is locked)

auphonic_review with only the short's audio clips and the Usual-5 preset (auphonic_presets; Polish on, de-esser 0.2), then auphonic_apply. The cleaned file is one clip on a new Auphonic track, not linked to the camera, and the originals are muted. Re-cutting later means cutting that clip too (the user splits it along), or cleaning again.

6. Verify

• frames_sheet over the short (16 frames) after each stage, plus frame_at on every highlight and logo moment.
• python3 ~/skills/short-edit-beacon/scripts/link_sync.py <project.beproj> after any ripple: a clip Ripple Delete only closes the gap on the tracks it touched, and once shifted a recording's audio 35 frames late from its camera on another track.
• Report: what fills each time range, the costs (B-roll dollars, motion graphics on OpenAI), and the spots worth a look.