openart-broll

OpenArt B-roll & Motion Graphics (MCP)

Everything here was proven in a shipped edit: three overlay clips, a reference-cloned talking head, and the billing/moderation scars to go with them.

Connection — this skill's one requirement

OpenArt has no API key. It runs over the OpenArt MCP connector (tools named openart_* appear in the session). Verify at the start:

openart_account_get → email, plan, credit balance. If the tools are absent, have the user connect the OpenArt MCP in their client (claude.ai connectors / Claude Code MCP settings), then re-check.

The call contract (every generation)

openart_model_list — pick the model whose description matches the job; modes matter: text2video (no reference), image2video (animate an exact first frame), element2video (subject/identity reference — pick this when unsure).
openart_model_form_get(model, mode) — exact fields (duration, aspectRatio, resolution, generateAudio, visualReferences…). Fill params to match.
openart_model_cost({model, mode, params}) — quote the exact config. Two billing quirks, both real: element2video quotes EXCLUDE a video-reference surcharge (a 5s clip quoted 300 credits billed 670 with one 14.5s reference); text2video bills at quote. Failed/moderated runs bill 0. The credit pool is shared with the user's web-UI activity — re-check openart_account_get before quoting affordability.
Show the user the verbatim prompt + config + credit estimate and get a yes BEFORE generating. Retries and "same config" re-rolls included.
openart_generate_video → historyId (PENDING). Poll openart_creation_get (respect pollAfterSeconds; a 5s clip ≈ 5–12 min). If the host shows a self-polling result card, don't poll — end the turn.
Download the result URL, re-encode dense keyframes (-g 30 -keyint_min 30), frame-extract and LOOK before placing it.

Recipe 1 — Motion-graphic overlays (per spoken line)

Generate short clips the editor layers over footage with a blend mode. The proven config: text2video, 4–5s, resolution matching the master, generateAudio: false, 9:16. The prompt contract — every clause earned its place:

flat solid black background (#000000) stated explicitly, plus "no background gradients, no vignette, no floor, no ambient glow filling the frame" — only BRIGHT elements (white/neon/brand color). Layered with Screen or Add blend, the black vanishes. Scope any glow "hugging the shapes only" (a blanket "no gradients" contradicts a requested glow and models drop one or the other).
One motion idea per clip, ending settled or as a seamless pulse (loop/freeze room for the editor). Name the clip duration in the prompt so "holds for the rest of the clip" means something.
Text ≤ 2–3 words, ALL CAPS, one string — renders cleanly even at 480p; more garbles. Zero-text icon clips are safest. Add "no letters, no numbers" to icon-only prompts or models invent labels.
State a placement band (top 20% / bottom 25%) but EXPECT it to be ignored — clips come back centered. Fine: it's a keyed floater; reposition the layer.
Append: "no watermark, no subtitles, no captions".
Verify the rendered background is truly flat black (frame-extract) — gray wash reads as haze under Screen blend.

Recipe 2 — Identity-referenced talking head / B-roll (`element2video`)

Give the model a reference video of the subject and direct new footage:

Getting references in: local files upload via the openart_upload_pick widget (the user picks the file; it can take minutes to register — poll openart_upload_list). Reuse prior uploads from openart_upload_list; openart_upload_metadata_get(url) mints a ready visualReferences object — including from a previous GENERATION's result URL.
Reference caps are enforced: video references must total ≤ 30s — a 30.04s clip is rejected at submit (input_invalid). Trim first.
Prompt pattern: address references by slot — The man in video 1 says: "<line>" ... — with camera/energy directions and "No background music, no captions, no on-screen text." Voice and identity clone from the reference.
Phonetic spelling beats mispronunciation: write acronyms as letters ("M C P"), dates as words ("July sixteenth"), and tricky names the way they sound. For non-English output, spell for THAT language's letter sounds.
Moderation trap (byte-plus models): the output-audio copyright detector can trip on the SPOKEN LINE itself (output_moderation_blocked / "output audio may be related to copyright restrictions"). No-music guard clauses do NOT fix it — reword the line. Failed runs bill 0, so it's cheap to iterate.
Whole-timeline b-roll beats follow the same grammar as arcads-broll: generate at the edit's aspect, slightly longer than the beat, match the grade, trim to the best 1.5–6s, QA hands/text/products, ~2 retries budget.

Placement

Same as any generated media: <video> in a composition card (unique id, data-start/data-duration), full-bleed takeover, or an overlay track with Screen/Add blend in HyperFrames, CapCut, or any NLE.