Broadcast News Package — Style Prompt

A television newsroom's on-air package, built entirely in code: a sting that becomes the channel bug, a rundown of today's stories that opens and closes the bulletin, a lower-third strap that never leaves the bottom of the frame, and full-frame graphics (split-flap price boards, a line breaking a threshold, explainer maps, documents flying out of a city, quote cards, spec cards with proportion bars, stamps) that carry the facts. The style is the package system plus its cutting rules, not a drawing style. 16:9, a clean frame: no burned-in subtitles, no news crawl. References (grammar only): evening network news headline menus, financial-channel market graphics, European news channels' explainer maps, tech-desk spec cards. Never copy a broadcaster's logo, set, fonts, colours, music or idents, and never name them in the film.

1. Essence, and what it is not

1. One package system, used consistently. Every graphic is built from the same few parts: a brand plate (parallelogram), a tab in the accent colour, a paper strap, ink cards with an accent top edge, mono datelines, stamps. Viewers read the system before they read the words.
2. The bulletin's order. Sting → rundown ("today we have four stories") → item → item → … → back to the rundown, ticked off → sign-off. Each item opens with its category on the strap's tab and its date on the first graphic.
3. One fact per screen. A full-frame graphic states one thing: the price, the break, the two stocks, the place, the two letters, the quote, the verdict.
4. The strap holds the headline. One lower-third strap is on screen from the first story to the recap and changes its words with a vertical flip; the picture above it is free for the graphic.
5. Clean panel. No crawl, no burned-in subtitles; subtitles ship as a separate .srt. The frame is the graphic, the strap, the bug and a source line.

Not Caption News (that is captions-first social video: huge slammed captions are the lead, no rundown, no strap; here type is small and orderly and the graphics lead). Not Data Storytelling (one chart per beat, never a dashboard; the chart is an insert in a bulletin). Not a keynote: no centred title-and-bullets, no page turns. Not a studio show: no virtual set, no anchor, no fly-throughs — the package is flat graphics over ink.

2. Frame and layout (1920×1080)

Content layers sit lower than a subtitled layout would put them (a 60–110 px downward shift) so the bottom strap does not leave an empty band above it. When you remove subtitles and the crawl, re-lay the page; never leave the space they used.

3. Package parts (specs)

4. Colour

Structure to keep when you change the brand: one accent + one gold-like secondary + paper + ink, and a reserved alarm colour. Ink text on the accent; on paper, English secondary lines in a dark olive (#6E5A1E) because gold is too weak there.

5. Type

• Chinese: Noto Sans SC 900 (headlines, strap) / 700 (labels).
• Latin: Barlow Condensed 800 (numbers, names, the plate), 600 (secondary lines, quotes), 600 italic (the end-card English title).
• Mono: IBM Plex Mono 500 for datelines, sources, axis labels, all caps with 1–4 px tracking.
• All OFL, subset from Google Fonts by demo/tools/fonts.sh (the character list is read from main.js; delete fonts/*.woff2 after a text change). No handwriting faces in this style.
• Every label is bilingual on one line each (Chinese first, English smaller); never wrap.

6. Picture layer and grade

• When there is no footage, the picture layer is data graphics drawn in code plus, at most, a few illustrative plates (generated or stock), each tagged ILLUSTRATION for as long as it is on screen. They set a mood (a server room behind a model launch, a wafer behind chip stocks); they never stand in for the event, the place or the people.
• One grade for every plate: greyscale → auto black/white points (0.5 / 99.6 percentile) → γ 1.05 → 35 % S-curve → black 8, white 232 → duotone #0E1013 → #ECE7D8 (tools/prep.py: grade()), plus a left-dark ink gradient for type, a 50–55 % vignette and 7 % overlay grain cycling through six tiles. Graphics and the package are never grained.
• Real footage, if you have it, takes the same grade path; modern footage ≤ 1.5× scale, archive ≤ 2.5× after deblocking; crop out station logos.

7. Editing grammar

• 120 BPM grid, beat 0 at the end of the sting (1.5 s). Every cut and wipe on a beat or half-beat.
• Shot lengths: sting 1.5 s; rundown 4–5 s; story shots 3–7 s; a map long take (6–12 s) as the bulletin's breath; a comparison stays until every item has been fully on screen ≥ 2.5 s.
• A package action every 0.5–1 s inside a shot (flip, count, stamp, tick, strap change, map light, line step). No frame flat for more than 1 s; a strong action within 0.5 s of every cut.
• Date every story: a weekly bulletin's items happen on different days; the date sits in the rundown row, the eyebrow line, the readout or the document.
• The order inside an item: what (the name or the number) → how big (the line, the bars, the price board) → who it hits or who said it (bars, documents, a quote) → the verdict stamp.

8. Sound palette

Master −14 LUFS, true peak ≤ −1 dBTP (core/render/mux.sh). Samples: VSCO 2 CE and VCSL (CC0); foley synthesised (core/audio/sfx.py).

9. Native moves

1. Sting becomes the bug — the brand plate slams in, then flies to the corner and stays there for the whole bulletin; opening and programme are one object.
2. Strap tab = category — the strap's tab changes category (模型 MODELS, 市場 MARKETS, 監管 POLICY …) with a white flash when a new story starts; no separate item cards between stories.
3. Rundown, then ticked off — the same list opens and closes the bulletin; at the end each row gets its tick.
4. Spec card — a launch reduced to its proportions: the whole and the part on one scale, then what it runs on, joined by a bracket to the total.
5. Strap that never leaves — one strap group from the first story to the recap; words flip, the strap itself never exits mid-programme.
6. Split-flap before/after — old and new value on one board, each tile flipping from the old to the new.
7. Breaking the threshold — a line advances one period per step towards a dashed threshold; the music stops for one beat before the last step and the crossing gets the heaviest hit.
8. Date in the eyebrow — every graphic's eyebrow line carries source and date in mono caps, so a weekly round-up never blurs which day each item happened.
9. City sends the letters — on a map long take, documents fly out of the sending city to their addressees; one world→screen function keeps labels and documents attached through push and pan.
10. Quote card + verdict stamp — the official's own words in type, with name, occasion and date; an underline on the key word as the voice says it; one stamp.

Where each move lives in the demo's code: DEMO.md (build notes).

10. Material

1. Facts first. Every on-screen number and every sentence of the voice comes from an official source (the company's own page, the exchange's data, the institution's press release or speech), recorded in demo/FACTS.md with claim, URL, quoted wording and any arithmetic. When a fact exists only in a news report, attribute it on screen and in the voice ("據…報道 / … reports") and keep to the plain claim; never put a single report's allegations about a real company or person on screen.
2. Data. Draw charts from real series only; when you have just a few reported numbers, draw a schematic line, tag it SCHEMATIC, and label only what you know. Exchange closes: the exchange itself, the central bank's statistics portal, or a market-data file whose source you log.
3. Pictures: ask in this order. a. The user's own images — photos, b-roll stills, their logo and colours (the package is rebuilt around their brand). Use these first. b. Generated illustrative plates, if the user wants them. Before generating, tell the user how many images and what each shows (one line per plate). With the user's Codex CLI signed in to an account that can generate images (core/image/gen_image.py --check), core/image/gen_image.py (driven by demo/tools/gen_plates.sh) makes them and saves JPEGs to demo/assets/plates/. With any other image tool (an API or a web app), give the user the prompts and the target file names; they generate the images and drop them into demo/assets/plates/. Prompts ask for no text, letters, logos, flags or people. c. No generated images: public-domain, CC0 or CC BY photographs (licences below), or abstract backgrounds drawn in code (gradients, grids, light fins — the package already carries the frame). Whatever the source: every word on screen is typeset in code, never baked into an image; a generated plate is only ever an illustration behind a graphic, tagged ILLUSTRATION / 示意 for as long as it is visible, and never stands in for the event, a place in the news, or a person. No generated anchor.
4. Photo and footage licences (judge each file on its own page, not the search result):
5. Wikimedia Commons: CC0, public domain or CC BY (credit author, title, licence on the end card and in CREDITS). Not used: CC BY-SA, "no known restrictions", unclear licences, YouTube re-uploads, colourised or "restored" versions.
6. U.S. federal agencies (NASA, NOAA, NARA, USGS…): public domain; check the page does not credit a contractor or third party.
7. Internet Archive: only items whose page states public domain, CC0 or CC BY; U.S. works published before 1931 are public domain.
8. Library of Congress / national archives: read each rights statement; "no known restrictions" is not enough.
9. Institutions' audiovisual services (for example the European Commission's): read their reuse terms per item; many require attribution and forbid implying endorsement.
10. Maps: Natural Earth (public domain), via world-atlas TopoJSON. Its default borders are not usable for China: DIRECTOR.md §13.
11. Before downloading, tell the user what you will fetch, from where and under which licence. The finder copied into this style, demo/tools/footage.py (search commons|archive|nasa, fetch, used, md), logs licence, author, page and the seconds used into SOURCES.jsonl and writes SOURCES.md. Keep raw/ under ~1.5 GB.
12. Logos: never use a company's logo for a company in the news; set its name in type.
13. People: public figures in the news can be named on a profile strap; no close-up or slow motion of private individuals.
14. Voice and music: ask whether the user has a recorded anchor, a preferred voice or a theme of their own before synthesising one.

11. Pitfalls

• Package left behind. If the strap, source line or tags are drawn once on top of everything, a wipe changes the picture under yesterday's headline and the strap catches up 0.3–0.5 s later. Draw them per picture, inside the wipe's clip.
• Leftover layout. Removing subtitles or a crawl without moving the content down leaves a dead band above the strap.
• Masked type peeking in. A push-in title clipped by a mask shows its top edge before it moves if the travel distance doesn't clear the mask; start it fully below and don't draw at progress 0.
• Words stronger than the source. "Request for information" is not "investigation" or "penalty"; write "no fine, no case opened" or "for the first time" only if a source says so — and check the institution's own pages, which can contradict a report ("first requests" may have gone out months earlier).
• Thresholds depend on the reference. "Bear market" = −20 % from which high? Name the reference on the axis; from a different peak the claim can be false.
• Attribute what only a report says. When the official source does not name the companies, the graphic carries "據… 報道 / reported by …" and so does the voice.
• A schematic posing as data. Rolling counters on a close price flash values that never existed; set the close with a mask push instead. Don't label axes the data doesn't support.
• Source lines under the strap. Right-aligned mono 13 px: ≤ 36 characters.
• Stamps over key words. A stamp hit can mask the voice ("fines"); take the brass off the stamp and low-pass its clack.
• Gold on paper has too little contrast: use ink or dark olive.
• Generated plates carry stray corner marks: crop them in prep.py.
• Short preview text. Rundown rows are previews of the stories; readcheck flags them as short in the opening — they return in the recap and as straps, so accept the flag or leave them out of TEXTS.
• Names in synthetic speech. Spell brand names, model numbers and abbreviations the way they should sound in the TTS text (a respelling, a number written as a word, an abbreviation without dots) while the strap and subtitles keep the real spelling; when speech-to-text mishears an unusual name, put the expected transcript in the line's asr field instead of rewriting the line.

12. Variation space

Decide per film: the channel's name, plate and accent colour; how many items (one lead with a long package, three or four equal items, a five-item round-up); which full-frame graphics each item needs; where the long take goes; the anchor's voice.

• Three structures: rundown → items → rundown ticked (round-up); one lead package with a correspondent two-box and a short "also today" grid; a countdown of items (election-night, sports results).
• Three openings: sting → rundown; cold open on the biggest graphic, then the sting; sting straight into the first story with the rundown as a recap only.
• Three endings: rundown ticked → end card; a "coming up next week" teaser card; the last stamp freezing into the end card.

How our demo was made (story, shots, score, end card, build): DEMO.md. Read it after your treatment exists.