kevinbadi/open-edits/fireship-style-short

Fireship-style SHORT (vertical faceless news short)

The long-form skill (fireship-style-edit) is the source of truth for the grammar: read its section 0 (LITERAL VISUALS) and section 3 (devices) first. This skill only changes what a phone needs. Both skills share one engine: fireship-style-edit/templates/fireship_engine.py, which switches to 1080x1920 when FIRESHIP_CANVAS=9:16 is set before import. templates/short_kit.py sets that and adds the vertical helpers, so a short's render.py starts with from short_kit import *.

Reference build: examples/weird-hack-short/render.py (the 60 s vertical cut of "The most interesting hack in history just got weirder", built from the same footage as the long-form edit).

1. What changes for vertical

2. Safe zones (platform UI covers these)

• Top 200 px: status bar, search, "Shorts" header. Stickers start at y ≥ 220.
• Bottom 460 px (y > 1460): title, channel, audio ticker. Captions sit ON the 1360 line (baseline),

nothing else goes below 1300.

• Right rail (x > 960, y > 900): like / comment / share buttons. Keep stickers left of it down there.
• The band's centre is BAND_Y = 760: stack a sticker ABOVE the band (y 220–420) and one BELOW it

(y 1080–1210, slot STICKER_BELOW = 1150). Never on top of the faces / lower third inside the band.

3. Pipeline

bash "${CLAUDE_SKILL_DIR}/scripts/setup_short.sh" <slug>        # projects/<slug>-short/ with engine + kit
cd projects/<slug>-short
# audio: the VO / overlay, trimmed to the short (30-60 s)
ffmpeg -t 60 -i "<audio>" -ar 48000 audio/voice.wav
# words: Whisper small.en + initial_prompt (names) -> audio/words.json   (same as long-form)
# footage: the long-form skill's fetch_footage.py (search -> download -> 1 fps sheet -> --cut [--crop])
python3 ../../.claude/skills/fireship-style-edit/scripts/fetch_footage.py --search "<person> <event>"

1. Shot list (time | exact VO words | what is literally on screen | source), ~1 shot per 1.5 s.
2. Assets, literal-first: footage of the named people/events, real docs (wikipedia_extract +

wiki_doc), real tweet cards (X mark + avatar), logos, memes (check burned-in caption position).

1. `render.py` from templates/render_short_template.py: Shot(...) list exactly like long-form,

but backgrounds come from footage_band / gif_band / image_band, cards go through card(), and render(SHOTS, DUR, post=make_captions(WORDS, KEYWORDS)).

1. Preview one frame per shot (PREVIEW=...), tile at 270x480, LOOK: captions clear of stickers,

nothing in the safe zones, faces/lower thirds in the band uncovered.

1. Render in 4 parallel RANGE chunks, short_mux('frames/out', 'audio/voice.wav', 'renders/final.mp4'),

copy to ~/Downloads/<Readable Name> (Short).mp4.

4. Clip + audio policy (Kevin 2026-10-02)

Any clip and any audio may be used in our edits (news footage, event footage, memes, a creator's audio overlay). Kevin's rule: no single third-party clip on screen for more than 10 s in a row. Cut away (sticker shot, diagram, meme) before 10 s and come back if needed. Don't add copyright caveats to deliveries. The continuous music bed is still owned / royalty-free, because it runs the whole video.

5. Don'ts

• No black or empty first frame. No static frame > 1.2 s.
• No sticker or card inside the safe zones, none over a face or a news lower third.
• No long-form-sized cards dropped in unscaled (they spill off a 1080-wide frame).
• No caption track fighting a sticker at the same height: stickers above the band or at 1080–1280,

captions on 1360.

• Everything in the long-form Don'ts still applies (literal visuals, real tweet cards, no em/en dashes in

sticker text, no Fireship branding).