PROJECT-SCOPED: this call acts only on the explicit project_id and returns the project identity with its result. Place a browser-rendered, After-Effects-grade motion graphic on the PROGRAM clock: spring/blur entrances, glow, gradients, 3D depth, masks, icon/emoji pops, counters, UI cards. One call = a finished composition, silent by default; pass sfx=true only for a moment that earns sound (its template cues map to the approved sound library and are levelled against the voice at their landings; CHECK lines flag a bright cue on a payoff word's onset or more than 1-2 sounds in a talking short). start/end are program seconds; end defaults to the template's natural duration. Cue it to the exact word/beat it amplifies (use word times from get_kept_transcript). layer='above_captions' (default for designed moments) or 'below_captions'. layer='behind_subject' is the premium depth signature: the graphic (a giant hero word, a number, a shape) is drawn INTO the shot and the speaker stays in FRONT of it — the person is measured out of the footage like add_text_behind (same rules: one continuous take, no speed ramp, >=0.4s; refused with the measured reason when there is no clear subject), and the reply reports how much of the graphic the subject crosses. Use it for 1-2 hero moments per short, with BIG type placed where the speaker's head/shoulders cross it. CAPTIONS: leave mute_captions unset — the captions drop exactly the spoken words this graphic shows (its number, slammed word, quoted kicker or rows) and keep every other spoken word (every heard word reaches the screen once — a lockup's rows no longer get micro bridge lines: the words they leave out are captioned beside it), moved to a band clear of the box it draws, the face and any card layout while it is up; the reply NOTEs the words captioned beside it and the words it must mute because no band is clear (then carry them on it, e.g. as a kicker, or move it off the caption band). mute_captions=true no longer hides a whole window; false keeps them all running (a number or starred word it shows is still not repeated). template='html' takes your own HTML/CSS/JS on the MG runtime in html. The write is rejected if the composition errors or draws nothing. HEADLINE BAND: template='headline' is PERSISTENT — the standing claim headline of a card or letterbox layout, one per program (end omitted = the program end; y omitted = centred in the free band above the card or picture). It yields its band on its own at render time (fades out before any graphic in the band lands, back after it leaves), never mutes captions and is silent; the reply lists the windows it yields. FACE KEEP-OUT: the write measures where the graphic draws against the speaker's face and mouth over its window (through the crop, zooms and cards) and the 9:16 safe area (6% side margins, the right 12% clear between y 0.5 and 0.85, nothing below y 0.80); a graphic that covers the face or leaves the safe area is MOVED to the nearest clear zone with its own y (and x/align/side where it has them) — above the head, below the chin, beside the face — and the reply says what moved (KEEP-OUT) or why nothing fits (NOTE). allow_face_overlap=true keeps a deliberate design over the face. NUMBERS land on their word: a counter completes 20 ms before its spoken number's onset in the transcript (the write sets its land; a count really counts, 0 -> value over ~0.4 s into the word, its window opening 0.45 s before it so it never runs across the setup; a 'reveal' starts on the word; a spoken range is value '30–40', each figure landing on its own word) and a word_slam whose hero is a figure ('32%', '$1.2B'; not a name like 'GPT-4') moves onto it — the reply says NUMBER LANDED, and NOTEs a moved window that now overlaps a neighbour. THE WINDOW: a word-timed graphic (phrase_build, marker_text over a spoken line) starts on its first visible word (WINDOW), and an entrance or exit within 0.15 s of a cut moves onto the cut frame so the change is one event (CUT-SNAP; never more than 0.15 s off its word; a number's entrance stays on its number). ONE TYPE SYSTEM: an accent/color you do not pass comes from the short's Look (the captions' highlight colour, else the accent its graphics wear; LOOK in the reply), and the reply NOTEs (look) a second accent, a fourth type role or broadcast furniture in an editorial Look, with the fix. phrase_build sets only its rows (at most 4, at most 3 sizes: SIZES); back-to-back slams of one style are a SERIES sharing one size per line. TIERS (word_slam tier, optional): 'hook' is the opening title as a headline (main line at 7%+ of the frame height, lead-in small above it) that owns its zone: the captions wait until it exits (HOOK; mute_captions=false keeps them); 'payoff' locks a number and its noun up in the accent ('140 / characters') — the short's largest lockup (NOTE (payoff) names a taller one); 'hero' is ONE giant word per short, behind the speaker by default on a measured person matte, drawn as a face-safe display slam above the picture when no person matte can be had (HERO FALLBACK). THE COVER (one call): template='cover_headline', start=0, end ~2.5-3, params {text: a strong curiosity title, 3-10 words, stars for the accent} sets the short's opening frame as a thumbnail: huge centred condensed caps, complete on frame 0, then it fades into the talk; it owns its zone like the hook tier (the captions wait until it exits; COVER). For SOME shorts, never every one.