# Skill: full/director — "במאי AI" (Complete video) The top of the range: the takes are cut and joined into one story, then Gemini watches and listens to the result — faces, pauses, what the words cannot say — and directs the whole edit: the hook, three generated clips on the moments the story shows (the one the customer asked for among them), kinetic words, punch-ins, flashes, sound design and an original score whose sections change exactly on the story's turns. One render, a finished video. **Includes:** takes combined into one video · Gemini watches and directs the edit · 3 AI clips on the story's moments · a dramatic score timed to the story's turns · sound design · kinetic words · punch-ins and flashes · hook · captions corrected by ear · colour grade · optional CTA end card and banner **Built on:** hook `punch` ("זום פאנץ׳", `/skills/hook-punch.md`) + captions `hormozi` ("הורמוזי", `/skills/captions-hormozi.md`) **Price:** 2150–2300 credits per finished minute · 30 s ≈ 2050–2150 credits **Uses:** `/v1/speech/transcripts` · `/v1/studio/trim` · `/v1/studio/merge` · `/v1/ai/text` · `/v1/ai/video` · `/v1/ai/music` · `/v1/ai/sound-effects` · `/v1/studio/edit` Base URL: `https://apihub.co.il` · Auth: `Authorization: Bearer ` · Docs: `/docs` Send the same `X-Tag: ` header on every call of one video; `GET /v1/usage?tag=` then returns exactly what it cost. Times in request bodies are **seconds on the edited (output) timeline**. Unknown fields are rejected with the name of the right field — read the error instead of guessing. **Files are kept 24 hours** (renders, uploads, generated clips/images/music) — download the finished video. **Models.** This template runs on the newest models: `quality: "best"` on every AI call — Veo 3.1 for clips, Nano Banana Pro for images, ElevenLabs v3 for Hebrew voice. Each is priced per asset, so a 30-second video costs almost as much as a 60-second one. `quality: "balanced"` is the cheaper cut of the same recipe. --- ## 1. Combine the takes (skip for a single take) The customer often sends a story and a re-take of the part that went wrong ("the pitch came out bad, so I shot it again"). Transcribe every take, find the cut points in the words, trim, then join whole clips. Retakes INSIDE one recording ("רגע, אני אגיד שוב…", a sentence said twice, "אז אני… אז אני רוצה", "את את") need no manual cut: add `"remove_retakes": true` to /v1/studio/edit (or short / remove-silence) — the engine finds them in the words, keeps the last complete take of each (`{ "mode": "best" }`: the cleanest one) and never cuts a word said only once; the render's `retakes` lists what it cut. Whole takes in separate files: merge them whole, then edit the merged video with `"remove_retakes": { "window_s": }` — or send the files to the video agent and pick "combine the best takes". ```jsonc POST /v1/studio/uploads { "filename": "take1.mp4", "content_type": "video/mp4", "size_bytes": } // → output.upload_url (PUT the file) + output.url · the ORIGINAL file as recorded, never a resized copy (a vertical video resized to 1920×1080 renders stretched) POST /v1/speech/transcripts { "audio_url": "", "language_code": "he", "speech_models": ["universal-3-5-pro", "universal-2"], "keyterms_prompt": [""] } // cut point = the silence between the last word you keep and the first word you drop: // "ברורה" ends 83.45 s, "אז" starts 84.36 s → cut at 83.9 (half a second of breath) POST /v1/studio/trim { "video_url": "", "start": 0, "end": 83.9, "output": { "resolution": "source", "quality": "high" } } POST /v1/studio/merge { "clips": [ { "url": "" }, { "url": "" } ], "output": { "resolution": "source", "quality": "high" } } POST /v1/speech/transcripts { "audio_url": "", "language_code": "he", … } → transcript_id (captions + timings) ``` Trim first and merge **whole** clips (no `start`/`end` inside the merge). Do not remove silences for a fast talker — the few seconds saved are not worth jump cuts in a story. For two versions (with / without a pitch) cut once and render twice: the shorter version is the same timeline trimmed. **Everything below is timed on the cut video's clock.** ## 2. Let Gemini watch the video and direct the edit Gemini watches the merged video (picture + sound) — it sees faces, gestures and pauses, hears what the words cannot say, and corrects the transcript. Give it the word timings too, for frame-accurate effects. ```jsonc GET /v1/studio/hook-styles · GET /v1/studio/caption-styles · GET /v1/studio/text-styles // the ids it may use POST /v1/ai/text { "quality": "best", // Gemini 3.1 Pro "system": "", "prompt": "Watch and listen to the whole video, then return the edit plan JSON.", "media": [ { "url": "", "type": "video", "fps": 2 } ], "json": true, "max_tokens": 32000, "temperature": 1 } // → output.json = the plan · ≈ 38 credits for 100 s of video · ~70 s ``` The director brief — send it as `system`, with the video length, the customer's must-have shots and, at the end, the word timings (`GET /v1/speech/transcripts/{id}` → `words`, ms ÷ 1000, as `word@seconds`): ```text You are the creative director and senior editor of a top short-form studio (Reels / TikTok / Shorts). You WATCH and LISTEN to the attached vertical video and design its edit. Be bold and creative, but every choice must serve the story — restraint is part of the craft. Return JSON only. THE VIDEO: s. Times below are seconds on this video's timeline. RETURN { "beats": [ { "start": s, "end": s, "beat": "…", "emotion": "…", "intensity": 1-10 } ], "corrections": [ { "at": s, "wrong": "word as transcribed", "right": "the word actually said" } ], "hook": { "text": "…", "style": one of ["box_red","yellow","black","outline","clean","type","punch","glitch","marker","words","box","counter","shake","rise","strike","blur","flash"], "duration": 2.5-3.5 }, "captions_style": one of ["classic","hormozi","karaoke","one-word","marker","news","neon","quiet","pop","typewriter","bounce","underline","elegant","handwritten","comic","sticker","drama","cinema","shadow","lime","retro","scale","bubble","sport","minimal","slide","volt","neon-glow","marker-tape","keyword-pop","keyword-box","editorial","poster","broadcast","soft","cards","cards-bold","cards-neon","kinetic-pop","fire","gold","glass","vhs","candy"], "grade": one of ["none","warm","cool","vivid","cinematic","teal-orange","bw","vintage","matte"], "clips": [ { "start": s, "duration": 4 | 6 | 8, "mode": "full" | "split", "prompt": "…", "why": "…" } ], "texts": [ { "start": s, "duration": 1.2-2.4, "text": "…", "effect": one of ["fade","pop","blur_in","slide_up","typewriter","highlight","punch","glitch","marker","words","box","counter","shake","rise","strike","flash"], "preset": one of ["hero","mega","accent","pill","pill-accent","marker","outline","small","script","type"], "y": 22-45 } ], "zooms": [ { "start": s, "end": s, "to": 1.05-1.3, "ease": "punch" | "smooth" } ], "flashes": [ { "start": s } ], "sounds": [ { "start": s, "prompt": "English, max 8 words", "duration": 0.5-3, "volume": 0.3-0.7 } ], "music": { "style": "overall sound, always instrumental, sits UNDER the voice", "sections": [ { "start": s, "end": s, "name": "…", "styles": ["…"], "avoid": ["…"] } ] }, "notes": "one paragraph: the idea of the edit" } RULES - Land every effect on the frame its word is said — take times from the word timings and from what you see. - beats: the emotional map from what you SEE and HEAR (voice, face, gestures, pauses). - corrections: speech-recognition mistakes you hear; keep slang and curses. - HOOK (0–3 s): the strongest line, 3–7 words in the video's language, never inventing facts; wrap ONE payoff word in *asterisks*. Prefer an effect style (punch, rise, words, glitch…) — the classic boxes cannot colour a word. - CLIPS: exactly 3, AI-generated, photoreal, cinematic, vertical 9:16: 1. MUST show: . Place it while it is being told, mode "full". The rest: the most visual moments of the story. Never over a punchline or the last line; at least 8 s apart; none before 3 s. prompt: English, 40–80 words: subject, action, setting, camera move, lens and light, mood; no real people, never the speaker; tasteful; end with "no text, no logos, no watermark". - TEXTS: 5–9 kinetic words, 1–4 words each taken from what is said at that moment, at most one per 5 s, never during the hook; counter for money and numbers; accent one word with *asterisks*. y = % from the top (captions live at 65–90). - ZOOMS: 4–8 — punch on surprises and contrasts, smooth on emotional lines; never during a "full" clip. - FLASHES: at most 3 (before a full clip or on a hard contrast). - SOUNDS: at most 12 — a whoosh 0.3 s before each clip, an impact on the hardest line, a hit on the turn, a riser into the payoff, soft pops on text effects. - MUSIC: consecutive sections covering 0 → exactly, each ≥ 3 s; borders land exactly on the story's turns (quiet tension in the set-up, a hit or drop on the conflict, a lift on the turn, a build to the payoff, a real ending). English styles: instruments, dynamics, tempo, energy. WORD TIMINGS (word@start seconds): <הם@0.00 יושבים@0.22 בפאב@0.48 …> ``` A real plan it returned for the example (84 s, trimmed): ```json { "hook": { "text": "הבחור שניגש לבחורה הכי *סקסית* בפאב", "style": "punch", "duration": 3 }, "captions_style": "hormozi", "grade": "cinematic", "corrections": [ { "at": 14.94, "wrong": "דפוסה", "right": "תפוסה" }, { "at": 70.71, "wrong": "מאמצים", "right": "אמיצים" } ], "clips": [ { "start": 7.6, "duration": 8, "mode": "full", "prompt": "An ordinary, slightly awkward guy approaches a stunningly attractive woman at a lively, softly lit pub. She politely turns him down, body language slightly closed off. Cinematic lighting, intimate atmosphere, shallow depth of field, natural colors. no text, no logos, no watermark.", "why": "ניגש לבחורה הכי סקסית בפאב … מצטערת, אני תפוסה" }, { "start": 77.2, "duration": 4, "mode": "split", "prompt": "A dull, monotonous office scene where uninspired workers sit passively at their desks in gray, dim lighting… no text, no logos, no watermark.", "why": "אנשים שיושבים בצד שקטים" } ], "texts": [ { "start": 32.61, "duration": 2, "text": "האומץ זה הניצחון", "effect": "blur_in", "preset": "hero", "y": 25 }, { "start": 52.31, "duration": 2.2, "text": "התגובה של *עצמנו*", "effect": "highlight", "preset": "marker", "y": 30 } ], "sounds": [ { "start": 14.8, "prompt": "deep bass impact on rejection", "duration": 1.5, "volume": 0.5 }, { "start": 22.7, "prompt": "vinyl record scratch wow moment", "duration": 0.8, "volume": 0.5 } ], "music": { "style": "Cinematic instrumental drama with rhythmic tension, evolving from subtle curiosity to a powerful emotional climax, always sitting under the voice.", "sections": [ { "start": 0, "end": 14.8, "name": "Pub Setup", "styles": [ "subtle tension", "curiosity", "minimal" ] }, { "start": 14.8, "end": 22.7, "name": "The Question", "styles": [ "suspense", "pause effect", "low drone" ] }, { "start": 22.7, "end": 42.2, "name": "The Twist", "styles": [ "uplifting", "driving rhythm", "hopeful" ] }, { "start": 42.2, "end": 60.1, "name": "Self Reflection", "styles": [ "serious", "introspective", "slow cello" ] }, { "start": 60.1, "end": 72.1, "name": "The Haters", "styles": [ "dark pulse", "edgy", "steady bass" ] }, { "start": 72.1, "end": 83.9, "name": "The Contrast Climax", "styles": [ "epic build", "staccato strings", "triumphant finish" ] } ] } } ``` **Your pass as the editor** before generating anything: keep clip 1 on the told moment (start it on the verb — "ניגש" — and let it run past the punch line it illustrates), make the most cinematic clip "full", drop any zoom that overlaps a full clip, reuse one generation for repeated sounds (three whooshes = one call). Veo renders 1080p only at 8 s — 4 s and 6 s clips are 720p. ## 3. Generate everything in parallel ```jsonc POST /v1/ai/video { "quality": "best", "prompt": "", "aspect_ratio": "9:16", "resolution": "1080p", "duration": "8", "negative_prompt": "text, captions, subtitles, watermark, logo" } // Veo 3.1 8 s ≈ 960 credits · ~110 s POST /v1/ai/video { …, "resolution": "720p", "duration": "4" } // 4 s ≈ 480 credits · ~50 s POST /v1/ai/music { "prompt": " Hybrid orchestral drama; instrumental.", "instrumental": true, "sections": [ { "name": "Pub Setup", "duration": 14.8, "styles": ["subtle tension", "curiosity"] }, … ], "avoid": ["vocals", "lyrics", "EDM drops"] } // durations = end − start; the track lasts exactly their sum · ≈ 0.75 credits/s · ~18 s POST /v1/ai/sound-effects { "text": "deep bass impact on rejection", "duration": 1.5 } // ≈ 1–2 credits each GET /v1/speech/transcripts/{id}/captions → the caption lines (ms); apply the plan's corrections to line and word text ``` ## 4. One render ```jsonc POST /v1/studio/edit // x-wait-ms: 110000, then poll GET /v1/executions/{id} { "video_url": "", "hook": { "text": plan.hook.text, "start": 0, "duration": plan.hook.duration, "style": plan.hook.style }, "captions": { "style": plan.captions_style, "captions": }, "overlays": [ { "url": "", "type": "video", "start": 7.6, "duration": 8, "mode": "full", "fade_ms": 180 }, { "url": "", "type": "video", "start": 77.2, "duration": 4, "mode": "split", "fade_ms": 180, "split_side": "top", "split_ratio": 0.55, "speaker_focus": 0.3 } ], "texts": [ { "text": "התגובה של *עצמנו*", "start": 52.31, "duration": 2.2, "effect": "highlight", "preset": "marker", "style": { "y": 30 } } ], "zoom": [ { "start": 22.7, "end": 28.4, "to": 1.25, "ease": "punch", "y": 0.35 } ], "panels": [ { "start": 22.74, "duration": 0.12, "color": "#FFFFFF", "fade_ms": 30 } ], // the flashes "sounds": [ { "url": "", "start": 14.8, "volume": 0.5 } ], "music": { "url": "", "volume": 0.26, "duck": false, "loop": false, "fade_in_ms": 400, "fade_out_ms": 250 }, "color": { "preset": plan.grade }, "audio": { "normalize": true }, "output": { "aspect_ratio": "9:16", "resolution": "1080", "fit": "crop", "quality": "high" } } ``` **Music level:** a talking head speaks almost all the time — with `duck: true` the score would be pushed under the voice for the whole video. `duck: false` at 0.26 keeps it ≈ 14 dB under the voice: heard, never in the way. The score was written to the story's turns, so its hits land by themselves. **Optional — a closing call to action and a banner** (the second version of the example): - End card: `POST /v1/ai/voice` (the pitch, ElevenLabs v3), a Veo 8 s background, `POST /v1/studio/edit` on it with `"speed": 0.88` (stretches 8 s to 9.1 s), `"audio": { "mute": true }` and the voice as `voiceover` → merge it after the story → in the final edit add its captions (transcribe the voice, offset the words), texts on it, and a dark panel (`"color": "#000000", "opacity": 0.55`) behind the last line so the URL reads. - Banner: one small text at the bottom, once in the middle — `{ "text": "נערך אוטומטית עם *API* *Hub*\napihub.co.il", "preset": "small", "effect": "slide_up", "exit": "fade", "style": { "y": 84, "size": 3.9, "weight": 800, "background": "#111111E6", "accent_color": "#FFD600", "max_chars_per_line": 40 } }`. ## 5. Check before you deliver Grab frames at the hook, inside every clip and on every text (`POST /v1/studio/thumbnail { "video_url", "at" }`, ≈ 1 credit each) and look at them: the accent word must be visible on its box, the captions clear of the texts, the clips on the right words. Re-render only what is wrong — a render of 84 s takes ~7 minutes. ## What the example cost and how long it took 84 s, two takes, three Veo 3.1 clips (8 s + 4 s + 4 s), a 6-section score, 8 sounds, one 1080p render: | Step | Credits | Time | |---|---|---| | Upload + transcribe 2 takes | 6 | 10 s | | Trim + merge | 186 | ~2 min | | Transcribe the merged video | 3 | 8 s | | Gemini watches and plans | 38 | 70 s | | 3 Veo 3.1 clips | 1,920 | ~2 min (parallel) | | Score (84 s, 6 sections) + 8 sounds | 72 | ~20 s (parallel) | | Final render (1080p) | 84 | ~7 min | | **Total** | **2,309** | **~13 min** | The clips are 83% of the price. The same recipe with Veo 3.1 Fast (`/skills/full-director-fast.md`) made the example for 749 credits. Send `x-wait-ms: 110000` to wait for the render in the same call; if the response is still `processing`, poll `GET /v1/executions/{id}` until `status` is `completed` and read `output.url`. ## Check the bill `GET /v1/usage?tag=` — per call, per operation and the total, in credits. Quote that to the customer, not the estimate above. ## Rules that keep captions readable on every platform - **Safe zone.** Keep text inside the middle 80% of the frame: `side_margin` ≥ 6 handles the sides; `margin` ≥ 12 keeps the bottom clear of the platform UI (username, buttons, progress bar). - **A line never leaves the frame** — a line too wide for the side margins is set smaller automatically. Do not shrink the size yourself, wrap instead (`max_chars_per_line`). - **Hebrew has no capitals** — keep `uppercase: false`; it only breaks Latin words mixed into a sentence. - **Fonts.** `GET /v1/studio/fonts` lists the bundled families and which scripts they draw. A font that cannot draw the caption's language is swapped and reported in `warnings`. - The customer's transcript wins: to fix a word, `GET /v1/speech/transcripts/{id}/captions`, edit the text, and send the list back as `captions.captions` — timing is kept when the word count is unchanged.