--- title: "Scripted Presenter Daily — talking-head segments, one call per beat" tier: hero format: video theme: presenter | daily | scripted persona: creator, marketer, comms duration: "~10–20 min" budget_usd: "~$6–$18 ($0.315 per second of speech — the script IS the bill)" caps: ["avatar-x-t2v", "avatar-x-ref2v", "ffmpeg-concat", "ffmpeg-burn-subtitles", "ffmpeg-export"] skills: ["avatar-x", "cinematography", "finishing-quality"] showcases: ["/chapters/scripted-presenter-daily-example.html"] status: "live" reliability: 3.6 # 5/5 across both endpoints at registration; slow (~316s) and priced per second of speech --- # Scripted Presenter Daily — agent playbook > Paste this whole file (with the BRIEF filled in) into Claude cowork, chat, or Code. > Livepeer Agent turns a script into a presenter segment — voice and performance in > one call, no separate TTS and no lip-sync pass — then cuts it for the platform. > > **Read the constraint first:** you choose an actor, and the actor's own setting > comes with them. The stock library is UGC/selfie-framed rooms, not sets — there > is no way to art-direct the background, framing or wardrobe. This playbook makes > creator-style pieces to camera, NOT a studio news desk. If you need a specific > location, generate the scene and put a talking head on it instead. ## What you'll get - **A presenter reading your script**, in the voice that comes with the avatar - **A second voice** for a guest beat, if the brief has one - **Burned captions** and a platform export - **A per-beat cost line**, because the script length *is* the bill ## Why one call, not three The usual route is `chatterbox-tts` → `talking-head`: generate the voice, then animate a still to it. Two billable steps, two failure points, and the lip-sync is only as good as the alignment. Avatar X does it in one: you give it words and an actor, it returns a person saying them. That is the entire reason to reach for it. ## Tell the agent about the segment ```yaml segment_title: # e.g. "Monday Update" beats: # one per shot. KEEP EACH UNDER ~40 WORDS — see the gates. - avatar: Jasmine script: "Good evening. Tonight on the desk: …" # 50–1500 chars - avatar: Michael script: "Thanks. From the engineering side, …" guest_audio_url: # OPTIONAL. A real VO track to drive a face instead of a script. guest_face_url: # OPTIONAL. The still or clip that VO should drive (ref2v). aspect: follows_avatar # NOT settable — each actor has a 16:9 variant, pick that captions: true max_spend_usd: 20.00 output_slug: # kebab-case filename prefix ``` ## Conventions (how the agent must run) 1. **`script`, never `prompt` — and no art direction at all.** Avatar X has no prompt field and rejects one outright (`extra_forbidden`). If the brief reads like a visual description ("a woman at a news desk"), say plainly that this capability chooses neither the set nor the framing: you pick an actor from 21 and get their room. Offer the alternative (generate the scene + `talking-head` / `veed-lipsync-v2`) rather than rendering something that will not match the brief. 2. **Length gate — the important one.** Output runs as long as the speech, billed at **$0.315/s**. Measured: **230 characters produced 13.48s of video** (~17 chars per second of speech) for **$4.25**, and took **~316 seconds to render**. State the per-beat dollar figure and the expected wait BEFORE rendering. Refuse to exceed `max_spend_usd`. Push back on any single beat over ~40 words: split it. 3. **Warn about the wait.** ~5 minutes per beat. Tell the user it is running; do not promise a time. 4. **Aspect is not a parameter.** It follows the avatar: bare names are 9:16 portrait, so choose the `(16:9)` variant when the deliverable is landscape. Passing `aspect_ratio` does nothing. 5. **No fallback.** If a beat fails, report it — do not substitute a different avatar or a stock clip. The user chose the presenter. ## Steps ### 1 — Render each beat ```jsonc run_capability({ capability: "avatar-x-t2v", inputs: { script: "", avatar: "" } }) ``` Show each beat as it lands, with its cost. **STOP** after the first beat for approval before rendering the rest — a wrong avatar or tone is cheaper to catch once than five times. ### 2 — The guest beat, if there is one When the brief supplies real audio (a recorded VO, a clip from an interview), drive a face with it instead of a script: ```jsonc run_capability({ capability: "avatar-x-ref2v", inputs: { audio_url: "", image_reference_url: "" } }) ``` Output length equals the audio's, and audio can be 180s upstream — at $0.315/s that is ~$56.70, so check the track's duration first and say the number. ### 3 — Assemble, caption, export ```jsonc create_media({ action: "generate", model_override: "ffmpeg-concat", inputs: { clips: ["", "", "…"] } }) create_media({ action: "generate", model_override: "ffmpeg-burn-subtitles", inputs: { video_url: "", granularity: "word" } }) create_media({ action: "generate", model_override: "ffmpeg-export", inputs: { video_url: "", width: 1080, height: 1920, mode: "cover" } }) ``` ## Deliver - the assembled segment, plus each beat on its own so a single line can be redone - **cost**: per beat and total, with the per-second rate so tomorrow's script can be priced before it is written - one honest line on whether the delivery suited the copy — this model reads a script, it does not act a brief ## When to use something else | If… | Use | |---|---| | you already have finished video and only need new audio | `veed-lipsync-v2` — ~4× cheaper, built for dubbing | | you have one still to make speak | `talking-head`, or `sync-lipsync-v3` for illustrated characters | | you need a specific real person's likeness | `heygen-twin` (15s of training footage) | | the segment is mostly b-roll with narration over it | generate the video, then `ffmpeg-mux` the VO — far cheaper | | the brief specifies a set, location or wardrobe | generate the scene, then `talking-head` / `veed-lipsync-v2` |