Stock photos and video

0. Library first

library list --kind sticker / --kind clip --query <words> may hold the picture (Online resources downloads too). A plugin may search stock: library search "clay pot" --kind sticker (a job); place with library place ID --at-frame F --size 0.6 --base-rev N.

1. Search

• All stock rows at once (plan.visuals rows needing stock): slots.json [{"slot": "pot", "query": "clay pot", "kind": "photo", "orientation": "portrait", "count": 8}, …], then uv run <skill_dir>/fetch_stock.py --slots slots.json --dest <project>/media/stock --max-mb <cap> downloads, saves to the library and draws numbered sheets. One look per sheet, then --pick "pot=3@120;river=2@300:90" places candidate n at a frame (:length); --help for single lookups. Ranges (tune): ≤ 30 s 4–8 slots × 6–8; 30–90 s 8–16 × 6–10; 1.5–5 min 16–40 × 6–10; longer 6–10 slots a minute × 5–8; video cap ~150 MB a minute.
• Sources (--source; default the first with a key, else free): pexels, pixabay (keys PEXELS_API_KEY, PIXABAY_API_KEY, environment only; the only vertical stock video; Pexels takes Vietnamese), free (Wikimedia + Openverse photos, licence per file), nasa (space, science; public domain), archive (Internet Archive photos, newsreels, home movies; by year, public domain), documentary (archive + Wikimedia + Openverse, news documentary only). No video key: say so, use stills with motion. Never place a Pixabay URL. Archive is slow (~0.3 MB/s): short items, trim to the moment.
• No key is no reason to skip: open https://www.pexels.com/search/<q>/?orientation=portrait (videos: /search/videos/<q>/) in a browser tool, read links and alt text, look at thumbnails before choosing. Photo ID = trailing number of /photo/<slug>-<id>/; download https://images.pexels.com/photos/<id>/pexels-photo-<id>.jpeg?auto=compress&w=1080.

2. Download

Into media/stock/ (photos ~100–250 KB at 1080 px; vertical video 15–60 MB); record nothing (--source URL only when at hand), the user handles rights. Avoid identifiable people. Reusable pictures go to the library:

bashcut library add --kind sticker --name "Clay pot" --file /abs/media/stock/x.jpg --tags food

3. Place

• No credits on the video or in its description unless the user asks; then import with --origin stock (web for a file from someone's own page, such as a CC BY clip) --license … --source URL --author "Name" (a second media import adds them to a file already in) and write the lines from project credits.
• Stock video: import, then media place on main (replaces a moment) or an overlay (illustrates one).
• Stills: media import /abs/media/stock/photo.jpg --place --track OVERLAY --at-frame F --base-rev N; placed for 3 s, trim to the words. Frame with transform (zoom < 1 for a card); a still that does not move looks frozen: clip motion ITEM --preset zoom-in (or pan-left/pan-right). Never convert photos to video files.
• Generated B-roll only when capabilities get library.generate --kind clip lists a provider (none is no reason to skip stock): library generate "<shot, framing, light>" --kind clip (paid: --dry-run first, say the price), library add --from-result JOB:N, library place. AI picture: the rules below apply, and project credits counts it in the AI share.

4. Rules

• No labels on the picture. Never write "minh hoạ", "stock", "AI", a source or a licence on the video: it makes the cut feel unnatural. Real footage stays the main picture; stock reads as illustration by context (over or between real shots, tied to the words), not by a caption. Never present stock as the real place in narration or titles.
• Tie each picture to its words ("nồi đất" → clay pot); 1.4–2 s each, a soft pop sound.
• How many: the stock and icon rows of plan.visuals; without one, 2–3 per video at most. Always off faces.

Icons (Iconify)

For a small illustration (money, pin, clock, check, arrow) an icon reads better than a photo. Iconify: 200k+ vector icons, free API, no key, ~0.2 s per call (tested October 2026):

curl -s "https://api.iconify.design/search?query=cash&limit=20"           # → icons ["mdi:cash", "solar:wallet-bold", …]
curl -s "https://api.iconify.design/collections?prefixes=mdi,solar"       # → each set's licence
curl -s -o cash.svg "https://api.iconify.design/mdi/cash.svg?color=%23FFD400&height=512"
rsvg-convert -w 512 -h 512 cash.svg -o cash.png    # transparent PNG; or: magick -background none -density 1536 cash.svg -resize 512x512 cash.png

• BashCut imports PNG, not SVG. Never convert with qlmanage: it paints a white background.
• Prefer filled sets (solid, fill, bold; fluent-emoji-flat for colour emoji): thin outlines vanish on a phone. ?color= recolours one-colour icons only.
• Save under media/stock/icons/, place like a still, 512 px, small (zoom ~0.25–0.4), with a pop; keep good ones as library stickers.

Emoji for keyword pops (Noto Animated, OpenMoji)

All pops in one call: uv run <skill_dir>/fetch_emoji.py --words "wow,money,fire,tiền" --dest <project>/media/emoji draws one sheet, a row per word. One look, then --pick "wow=1@120:30;money=2@300:30" saves and places each. Land it on the payoff word, off faces, with a pop; length from bc:effects.

Brand and tool logos (Simple Icons)

uv run <skill_dir>/fetch_logos.py --brands "figma@120:90,notion@210:90" --dest <project>/media/logos --on dark renders, saves and places each in the brand colour (white/black under 3:1 contrast); no look needed. A near miss prints close slugs: rerun with the slug. Logos are trademarks: only to name the product shown or reviewed, never to suggest sponsorship, endorsement or partnership, never recoloured to the channel's palette or altered.

Illustrations in one style (open sets)

For explainers and promos an open set beats image.generate: no cost, one style. Generate only what no set draws (a named place, a product, a mascot). Sets: undraw (~1,800 flat topic scenes, one accent; the widest for explainers), humaaans (90 flat people), open-peeps (93 hand-drawn people with faces), open-doodles (33 sketchy scenes). One set per video, never mixed (the first --set is stored).

slots.txt (slot: words lines), then uv run <skill_dir>/fetch_illustrations.py --slots slots.txt --dest <project>/media/illustrations --set undraw --palette "#accent,#second" (brand palette, else two plan colours) draws numbered sheets. One look per sheet, then --pick "hook=1@0:90;idea=3@300:75" places a transparent PNG; a bare n prints its ID and file (HyperFrames input). Never write or read SVG.

Generated pictures (an image.generate plugin): set the style once, bashcut image style --style "<style>" --references <character.png>, then all scenes in one job, bashcut image generate prompts.json (scene prompts only; the lock adds the style). One look at its contact sheet, re-generate only failed or off-model scenes, media place. No AI label on the video (it looks unnatural): no --disclosure, no image disclose unless the user asks; AI disclosure is the platform's upload toggle (bashcut.vlog:publish). Mascot poses: bashcut library brand-pose <kit> --pose thinking --image photo.jpg.

Free fallback: blurred own footage

For a background (horizontal clip in a vertical frame, split moment, info card): a copy of the clip on the layer below, zoomed to fill, darkened and softened with an adjustment until the subject reads. No value fits all; as an example only, zoom ~1.8, exposure −1, saturation 0.7 worked; a bright, busy background needs more, a dark one less. Check with ui frame F --phone. Same colours, no licence question.