Two kinds of generated material for an edit:
- B-roll beats — AI clips that replace the talking head for 1.5–6s (product shots, scene beats, UGC inserts, stylized illustrations of a spoken claim).
- Motion-graphic overlays — short generated clips designed to be layered OVER footage with a blend mode, when you want a generated look instead of (or on top of) hand-authored composition cards.
Dependency: the Arcads skill pack
The API mechanics (auth, routes, model prompt library, polling, cost logging) live in the companion pack — install it once alongside this repo:
git clone https://github.com/krusemediallc/arcads-claude-code
Then follow ITS arcads-external-api skill for every call: base URL
https://external-api.arcads.ai, HTTP Basic auth with ARCADS_API_KEY as
username / empty password (put the key in THIS repo's .env too so one env file
serves the session), POST /v2/videos/generate for video models,
POST /v1/b-roll and POST /v1/scene for product-level b-roll,
POST /v2/images/generate for stills. Read that pack's
prompting/prompt-library/<model>.md before composing any prompt — its per-model
formulas are the difference between usable takes and slop. No account? Sign up at
https://arcads.ai/?via=claude-code.
Cost gate (inherited rule): estimate credits BEFORE generating (that pack's logs → its MASTER_CONTEXT rates → ask), present the estimate, and wait for a yes.
Editor-specific rules (this repo's additions)
Generating B-ROLL BEATS for a cut
- Generate at the EDIT's aspect (9:16 for reels) and slightly LONGER than the beat window — you trim to the best 1.5–6s, never stretch.
- Match the grade direction of the surrounding footage in the prompt (warm/neutral, contrast level) or plan a color pass; mismatched grade is what makes AI b-roll feel pasted in.
- Follow the reference-reel grammar if a STYLE-GUIDE.md exists: b-roll replaces the face for proof/drama beats, 1.3–4.1s, never longer.
- Before placing: re-encode dense keyframes (
-g 30 -keyint_min 30), then frame-extract and LOOK — check hands, text, logos, warped products. Regenerate defects (budget ~2 retries), don't ship them. - Stills for shot-cards (Nano Banana / gpt-image-2) display like real screenshots:
sharp card over a blurred blow-up of itself, or inside a browser-chrome frame
(branded-ad-edit
references/design.md).
Generating MOTION-GRAPHIC OVERLAYS (the screen-blend trick)
Generated overlays composite cleanly when you control the background. Proven prompt contract — every clause matters:
- flat solid black background (#000000), no gradients/vignette/floor, and only BRIGHT elements (white/neon) → layer with Screen or Add blend and the black vanishes. (True alpha isn't a thing text-to-video models give you; black+Screen is the portable substitute.)
- One motion idea per clip (a slide-in, a count-up, a pulse); end settled or seamlessly looping so the editor can freeze/loop to fill the beat.
- On-screen text ≤ 3 words, ALL CAPS, one string — more garbles. Zero-text icon clips are safest.
- State the placement zone (top 20% / bottom 25%) in the prompt BUT expect models to ignore it and center the graphic — that's fine: it's a keyed floater, you reposition the layer in the composition or editor.
- Append: "no watermark, no subtitles, no captions".
- Verify each clip's background is truly flat black before compositing (frame-extract; a gray wash under Screen blend reads as haze).
Placement
Generated beats drop into the composition like any media: a <video> inside a
card (unique id, data-start/data-duration, dense keyframes), or full-bleed via
a takeover card. Overlay clips go on a track above footage with the blend mode —
in HyperFrames, CapCut, or any NLE.
Provenance
Distilled from shipped edits that used generated overlay packs (flat-black Screen-blend motion graphics with word-synced pops) and the Arcads pack's b-roll / scene / UGC flows. If your own past project encoded further moves you want here, point the session at it and extend this file.