---
name: ugc-realistic-actor
description: "A creator so real nobody can tell. Paste your script, pick the vibe, get an authentic talking-to-camera UGC video, in any language."
---

# UGC Talking Actor

Goal: one talking-actor UGC video that is impossible to tell apart from a real person filming on their phone. The test for every choice: "would a real person, in this exact place and situation, film it this way?" Realism comes before everything else: skin, how the phone is really held and what the arm rests on, camera motion with a physical cause, a specific real place, real-life styling, emotion that fits each line, real sound. **The actor says the script word for word: no added words, no fillers, no missing lines.** Deliver exactly ONE final video.

**Scope: a person talking to camera, nothing else.** This skill never shows a product. If the user wants a product in the video (held, used, shown), say in one line that this skill only makes talking-actor videos and point them to a product skill (e.g. `cinematic-ads-ecom`, `cirkul-ugc` for Cirkul). Don't add the product yourself.

**A phone is never shown.** The actor is always filming themselves, so the phone is the camera: no phone in the frame, no second phone in a hand, no mirror selfies, no propped or mounted phone. Every video is a handheld front-camera selfie, and the prompt never gives the phone a hand, a position or an action (step 2a, "Phone wording").

All generation runs through the Arcads MCP. Load tool schemas with ToolSearch before calling them (`arcads_generate_video_flux_3_max`, `arcads_generate_video_seedance_25`, `arcads_generate_image_gpt`, `arcads_extend_video`, `arcads_extract_frame`, `arcads_stitch_videos`, `arcads_get_asset`, `arcads_view_frame`, `arcads_get_credits`, `arcads_get_upload_url` for user files).

## Benchmark prompt (the reference for every prompt this skill writes)

Tested Nov 2026, FLUX 3 Max, text only, 20 s, French, 73 syllables: word-perfect, rated the best result so far by the team. Every prompt follows its shape: short natural prose (~350 words), a specific room, a handheld rig with a physical cause for the motion, then the dialogue as timestamped beats, each carrying its own camera move and emotion.

> Authentic vertical UGC selfie video filmed on an iPhone front camera, one continuous take, no cuts. A French woman around 28, messy shoulder-length light-brown hair, no makeup, black oversized hoodie and grey sweatpants, sits on an unmade bed with white sheets, knees pulled up, in a bright Parisian Haussmann bedroom: tall window with wrought-iron balcony, hanging plant, bedside table with a small lamp, cast-iron radiator. Soft natural daylight from the window. She holds the phone in her left hand with her elbow resting on her raised knee, so the phone sits low, about 40 cm from her face, filming slightly upward at her (gentle low angle, like a real casual selfie). Framing: head and shoulders, top of the hoodie, the ceiling and window visible behind her. Her arm is relaxed, never stretched out. The camera rocks slightly as her knee and body move: organic micro-shake, small tilts, gentle sway.
> - 0-5s: she rests there, a bit blasé: "Franchement, je pensais que c'était normal que ça sente quand t'as un chat."
> - 5-8s: she lifts her elbow off her knee and brings the phone up close to her face, eye level, tight close-up, like a confidence: "Et puis ma copine m'a fait tester Pacha."
> - 8-15s: she settles the elbow back on her knee, the frame drops back to the low angle with a small wobble; she gestures with her free right hand and brightens: "C'est des petits cristaux qui piègent l'odeur, et ça arrive chez moi chaque mois, plus de sacs à porter."
> - 15-20s: she tucks a strand of hair behind her ear, smiles, then points down: "Franchement, je regrette de pas l'avoir su avant. Le lien est juste en dessous."
>
> She speaks in French with a natural Parisian accent, fast conversational pace, like a FaceTime to a friend. Audio: only her voice recorded by the phone mic, light room tone, no music. Realistic skin texture, front-camera softness, auto-exposure shifts as the phone moves.

What made it work (keep all of it):
- **Handheld, phone never seen:** the benchmark got away with "She holds the phone in her left hand…", but later takes showed that any sentence giving the phone a hand, a position or an action can make FLUX draw it. The skill now never does: it describes the selfie point of view and what the frame does (step 2a, "Phone wording"). Propped phones and mirror selfies are never used.
- **Rig with a physical cause:** the arm rests on something (elbow on raised knee), so the camera moves because her body moves. Never shake without a cause.
- **Low, close, relaxed:** about 40 cm, gentle low angle, arm relaxed, never stretched out.
- **Timestamped beats:** each line has its time window, its body action, its camera beat and its emotion, so the model paces the speech to the clip.
- **One motivated push-in:** a body action brings the frame close "like a confidence" for a whole line, then the frame drops back with a small wobble.
- **A specific real place:** 3–5 concrete objects (window type, plant, lamp, radiator), one real light source.
- **Short and positive:** no walls of "no X" negatives; the only negatives are "no cuts" and "no music".
- **Duration matched to the script:** 73 syllables at ~3.6 syllables/s ≈ 20 s, so the model had no spare second to fill (step 4 adds 1.5 s for the ending and rounds up, capped at the model maximum).

## 0. Model choice (decided by the skill, never mixed)

Tested winners: FLUX 3 Max is the most realistic from text alone; Seedance 2.5 is the most realistic with a reference image; GPT Image 2.5 Sunburst makes the most realistic actor-in-scene images.

**Language rule (checked FIRST):** if the script is in **any language other than English**, the first clip is **always FLUX 3 Max** (up to 20 s), whatever the actor source; anything beyond 20 s is added with **more FLUX 3 Max clips, each starting from the previous clip's last frame**. Seedance 2.5 and Kling 4.0 mispronounce non-English words and numbers (tested in Dutch, Nov 2026: same prompt, same 10 s script; FLUX 3 Max word-perfect with a native accent; Seedance said "zevenendertig" as "zevenhonderddertig" and garbled "betaalt" and "stroom"; Kling said "zevenendartig" and "Dat's"; Omni Flash was word-perfect but capped at 10 s and ignored the styling). Never use Seedance or Kling for a non-English script. Never use Seedance, Kling or Omni Flash to extend a non-English video either: FLUX chained from the last frame is the only way.

### Non-English path (FLUX 3 Max only, chained from the last frame)

| # | Actor source | Length | Clip 1 (FLUX 3 Max) | Beyond 20 s |
|---|---|---|---|---|
| N1 | Generate natively | ≤ 20 s | Text only, the whole script | — |
| N2 | Generate natively | > 20 s | Text only, first ≤ 20 s of the script | FLUX clips from the last frame, ≤ 20 s each |
| N3 | Image for my approval | Any | Approved Sunburst image as start frame | FLUX clips from the last frame, ≤ 20 s each |
| N4 | Use my image | Any | User image (or its approved Sunburst version) as start frame | FLUX clips from the last frame, ≤ 20 s each |

- **FLUX takes no reference images, only a start frame.** A specific actor (Sunburst or user image) goes into the clip 1 start frame.
- **Longer than 20 s:** clip 1 = FLUX 3 Max with the first ≤ 20 s of the script (cut at a sentence end). Each remaining part (≤ 20 s, cut at sentence ends) is a new FLUX 3 Max clip: `arcads_extract_frame` with `frame: "last"` on the previous clip, then `arcads_generate_video_flux_3_max` with that frame as `startFrame` and its own `duration` (step 4). When every clip is done, `arcads_stitch_videos` [clip 1, clip 2, …] in order. One clip at a time, never in parallel (each needs the previous clip's last frame).
- **Every new part starts with a camera change** (step 2a, "Cuts between parts"), so the join looks like a real creator's jump cut instead of the same video glitching.
- Every part uses the same actor, styling, scene, rig, voice and audio wording, word for word.
- Phone in frame: FLUX draws the filming phone whenever the prompt makes it an object in the scene. Tested Nov 2026: "phone propped against the fruit bowl" (kitchen) showed the phone twice, even when rewritten as "The camera is the phone propped on the counter…" with the POV line; "holds the phone in his right hand" while walking showed it once; "The camera is the phone in his right hand…" + "the phone itself is never visible" fixed one street take but still showed the phone in a park walking take. So: always handheld and always the "Phone wording" rule (step 2a): the phone never gets a hand, a position or an action in the prompt, and no line names it to forbid it.

### English path

| # | Actor source | Length | Video model | References passed to the model |
|---|---|---|---|---|
| 1 | Generate natively | ≤ 20 s | **FLUX 3 Max**, 1 clip | None (text only) |
| 2 | Generate natively | 20–30 s | **Seedance 2.5**, 1 clip | None (text only) |
| 3 | Generate natively | > 30 s | **Seedance 2.5** clip 1 (≤ 30 s) + **Seedance 2.5 extends** | Clip 1: none. Extends: run on the latest video |
| 4 | Image for my approval | Any | **Seedance 2.5** (+ extends beyond 30 s) | Sunburst actor image on clip 1 and on every extend |
| 5 | Use my image | Any | **Seedance 2.5** (+ extends beyond 30 s) | User actor and/or location image (or its Sunburst version) on clip 1 and on every extend |

Rules (both paths):
- **"Generate natively" is always the recommended actor source**, long or multi-clip videos included.
- **One model per path.** English: never mix FLUX and Seedance, never chain FLUX clips; beyond 30 s, only Seedance extends. Non-English: FLUX only, every clip after the first starting from the previous clip's last frame.
- The 20 s and 30 s limits are the **clip length** (speech time + the 1.5 s ending, step 4), not the word count.
- **Any English video over 30 s:** clip 1 = Seedance 2.5 with the first ≤ 30 s of the script (cut at a sentence end). Every remaining part (4–30 s of speech, cut at sentence ends) is added with `arcads_extend_video`, `model: "seedance_25"`, on the latest video, with the same reference images as clip 1. **The result already contains the whole video, stitched**: no trim, no stitch. One extend at a time. Cost: an extend also bills part of the source it builds on (13 s of new footage cost 2,162 credits in the Nov 2026 test), so it costs more than the new seconds alone.

## 0b. Credit estimate (always shown before generating)

Before generating, tell the user the estimated credits: every part and the total in the plan (one "Go?" covers the whole pipeline), and each regeneration on its own. Format, e.g. "Next: FLUX 3 Max, 1080p, 14 s → ~980 credits." Never launch a paid generation whose cost the user hasn't seen.

Rates observed on Arcads at 1080p (credits = seconds × rate, round up):

| Tool | Rate |
|---|---|
| FLUX 3 Max, 1080p, audio | 70 credits / second |
| Seedance 2.5, 1080p, audio | 104 credits / second |
| GPT Image 2.5 Sunburst | 0 credits within the daily image allowance (else report what Arcads charges) |
| Seedance 2.5 extend, 1080p | new seconds × 104 + part of the source (13 s new on an 11 s source = 2,162); estimate ~1.6× the new seconds, read the real charge |
| Extract frame, stitch | 0 credits observed |

- In the plan message (step 4), show a cost line per clip and the **total for the whole video**, and the current balance from `arcads_get_credits` if available. Then ask once: "Go?" (one question) and wait for the yes before generating.
- Before a regeneration (failed clip), state its cost again and ask.
- After generation, read `creditsCharged` from `arcads_get_asset` and report the actual total at delivery. If Arcads charges differently from the table, trust Arcads and say so.
- If Arcads returns `PRODUCT_SELECTION_REQUIRED` (it files assets under a brand), ask which brand to file under (one question) and reuse it for every call.
- If a generation fails with a download / "object does not exist" error twice on the same prompt, the provider is likely silently blocking the output: remove real place names and body-fluid or exertion wording (sweat, out of breath) before retrying (tested Nov 2026: a Prospect Park post-run prompt failed twice; a generic park walk with neither passed).

## 1. Intake: ONE question per message

Ask exactly one question per message. Wait for the answer before asking the next one. Never group questions, never list upcoming questions. Keep each question to one or two short lines.

1. **Script**: "What's the script? Paste the exact words the actor says, or I can write one." If Claude writes it, follow the script rules in step 1b.
2. **Language**: ask only if it isn't obvious from the script. Any language other than English → non-English path (FLUX 3 Max, chained from the last frame beyond 20 s).
3. **Format** (mandatory, never skipped): "Which format: 9:16, 16:9, 1:1 or 4:5?" Always ask it as its own question with AskUserQuestion (9:16 first, marked Recommended). Never assume the format from words like TikTok, Reels or vertical; skip it only if the user wrote an exact ratio (e.g. "9:16").
4. **Actor source**: estimate the clip length from the script first (step 4), then ask with AskUserQuestion, three options in this exact order, each with its estimated credits (step 0b):
   - **Generate natively in the video (Recommended)**: the video model creates the actor from a description. English: FLUX 3 Max if ≤ 20 s, otherwise Seedance 2.5 (extended beyond 30 s). Non-English: FLUX 3 Max, more FLUX clips from the last frame beyond 20 s.
   - **Generate an image for my approval**: a Sunburst actor-in-scene image first, you approve it, then Seedance 2.5 (English) or FLUX 3 Max with it as start frame (non-English).
   - **Use my image**: you attach an actor and/or location photo, then Seedance 2.5 (English) or FLUX 3 Max with it as start frame (non-English).

   **Native is ALWAYS the first option and ALWAYS the only one labelled "(Recommended)"**, whatever the length, the number of clips or the scene. Never present the image options as better. Consistency is never a reason to recommend an image: on native multi-clip videos, every part after the first continues the video so far (Seedance 2.5 extend in English, a FLUX clip from the last frame in non-English).
5. Depending on the answer:
   - Use my image → "Attach your actor and/or location photo." (if an actor photo is given, skip the actor description)
   - Native or image for approval → "Describe the actor: gender, age, look, hair, vibe (e.g. 'tired mom, 35')." If the user says "decide", pick one that fits the script and brand.
6. **Location**: "Where is the actor filming? (car, bedroom, kitchen, bathroom, street, office, gym, club…)" Skip if a location photo or the request makes it obvious.
7. **Time of day**: ask only if it changes the scene (e.g. car, street, bedroom). Skip otherwise, or decide if the user said "decide".

Format vs model: FLUX accepts 9:16, 16:9, 1:1, 3:4 (use 3:4 for 4:5). Seedance accepts only 9:16, 16:9, 4:3. If the chosen path can't take the requested format, say so in one line and propose 9:16 (one question).

Don't start generating until all answers are in. Only the actor talks to camera.

## 1b. Script rules (when Claude writes or cleans the script)

- **Brand names, URLs, acronyms and numbers are written the way they are spoken, in the script's language**: "meilleurtaux.com" → "meilleur taux point com"; "arcads.ai" → "arcads point a i"; "37%" → "zevenendertig procent". Never leave a raw URL or digit in the dialogue: the model reads it letter by letter or in the wrong language. Common acronyms the model already says right (e.g. French "l'IA") stay as written, with no extra pronunciation hint: a hint made FLUX say the word twice.
- **No colons, ellipses (…), dashes or parentheses** in the dialogue: they make the model stumble or improvise. Use periods and commas only.
- Short, simple sentences, 4–5 max for a ≤ 20 s clip.
- A user's "X sec max" is a ceiling, never the target: the clip length always comes from the script (step 4).
- If the user pasted their own script, keep their wording; only propose the phonetic spelling and punctuation cleanup in one line.
- When Claude writes the script, use the structure of the benchmark ad (≈ 20 s, 4 beats): a confession hook ("Franchement, je pensais que c'était normal que…"), how she found out through a person, ending on the brand name ("Et puis ma copine m'a fait tester Pacha."), two or three concrete benefits in one spoken sentence, an emotional line ("Franchement, je regrette de pas l'avoir su avant."), a soft CTA ("Le lien est juste en dessous."). First person, spoken words, no ad vocabulary. Match grammatical gender to the actor ("crevée" for a woman, "crevé" for a man).

## 2. Creative direction (decided by the skill, not asked)

From the script and location, decide these yourself and write them down before generating. Show them to the user in a short list together with the model choice and the clip table (step 4).

### Real-filming brief (always first, for every video)

Before writing anything, picture the real person behind this video and answer these questions for THIS location, THIS time of day and THIS script. The answers become the prompt. The tables and examples below are starting points, never a substitute for this reasoning: a bedroom at 8 am, a bedroom at midnight and a dorm room are three different videos.

1. **Situation:** why is she filming right now, and what was she doing just before (woke up, back from work, waiting in the car, between sets, out with friends)? Her energy, clothes and position follow from that.
2. **Body position:** sitting, lying, standing, walking, leaning? On what exactly (bed, car seat, counter, kerb)?
3. **Phone hold (always handheld):** which arm films, what her arm or elbow rests on (knee, armrest, counter, table, nothing), distance in cm. A real person rests the arm whenever they can and only stretches it when they have to.
4. **Angle and framing:** the angle that hold gives naturally (low from an elbow on a knee or a counter, eye level while walking, high only if she is lying down), how much of her body is in frame, what is visible behind her.
5. **Motion and shake:** what physically moves the frame (her knee, her steps, the car seat, her weight shifting, the music, her breathing) and how strong it is. Calm places: small sway. Lively places: real jolts. The intensity matches the place, never the same everywhere.
6. **Movement over the clip:** how her position changes with the script (settles in, the frame pushes in close on the key line, sits back on the payoff). One motivated push-in at most.
7. **Gestures and actions:** what her free hand naturally does there (hair, coffee, steering wheel, points down), 1–2 per 20 s, tied to a word.
8. **Light:** every real source at that hour (window, lamp, street lights, dashboard, club lights, screen glow on her face), its direction on her face, and how auto-exposure reacts when she moves.
9. **Voice:** volume and pace a real person would use there (soft late at night, normal at home, louder in a café, shouting in a club), mouth-to-mic distance, native accent of the country.
10. **Sound:** the room's acoustics (dry bedroom, closed car, open street) and at most 1–2 real background sounds of that place at that hour, at their real level. Keep the voice close and dry indoors: a tiled-kitchen echo plus a hissing moka pot came with garbled French on one take (Nov 2026, unconfirmed).
11. **Cuts:** one continuous take inside a clip, like a real selfie. Only between parts of a long video: a jump cut with a small reframe, like a creator restarting the recording.
12. **Local details:** 3–5 objects that make the place real and match the country of the script. Describe places generically ("a quiet city park", "a Brooklyn bedroom"), never a named landmark.

**AI tells to remove:** studio or beauty lighting, too-clean audio, a perfectly centered and level frame, uniform jitter with no cause, a model-like actor, perfect hair and skin, an empty or frozen background in a busy place, staged gestures, any phone visible in the shot, a static tripod-like frame, music or sound effects that the place wouldn't have.

### Core principle: be IN the action

The actor is doing what real people do in that place, right in the middle of it. Never move them to a quiet corner, a hallway or a calm spot to make the shot easier. Club → in the crowd, dancing. Gym → between sets. Street → walking through it. The energy of the motion, the camera and the voice matches the energy of the place.

### a. Phone rig and motion: chosen from the location

Every rig is a **handheld front-camera selfie**. Decide **which arm films (and is out of frame), what that elbow rests on (or how the arm is bent), the distance in cm (relaxed arm, never stretched out), the angle, the framing (with what's visible behind her), and the physical cause of the frame's motion**. The motion words ("organic micro-shake, small tilts, gentle sway") always come right after their cause, never alone.

**The phone is the camera, never an object in the shot.** Never a propped, mounted or tripod phone, never a mirror selfie, never a second phone: each of these makes the model draw a phone (tested Nov 2026: a phone propped on a kitchen counter showed up in the frame on two takes).

**Phone wording (every prompt).** Any sentence that gives the phone a hand, a position or an action makes FLUX picture someone holding a phone and film it from a second camera (tested Nov 2026: "the camera is the phone in his right hand, arm bent…" + "he brings the camera up close" + "the phone itself is never visible" still showed the phone in a walking park take). So:
- The word "phone" appears only in the first sentence ("selfie video filmed on an iPhone front camera") and in the audio line ("phone mic"). The word "camera" appears only there too. Never "the phone", "the camera" or "her phone hand" anywhere else.
- Write the rig as a point of view: "Filmed from her own selfie point of view, about 40 cm from her face, slightly from below, her left elbow resting on her raised knee. Her left arm is out of frame. The only hand that ever enters the frame is her empty right hand."
- Write every movement as the frame moving: "the frame rocks as her knee moves", "the frame pushes in close to her face", "the frame drops back with a small wobble". Never "she brings the phone / camera closer".
- No line that names the phone to forbid it ("the phone is never visible"): it names the phone again.

The table's rig column is the reasoning; always translate it into this wording in the prompt.

| Location | Rig (handheld) | Cause of motion | Motion level |
|---|---|---|---|
| Bed / couch | Sitting with knees pulled up, elbow resting on her raised knee, about 40 cm from her face, slightly from below. Head and shoulders, ceiling and window behind her | The frame rocks as her knee and body move | Low-medium |
| Bed at night | Lying on her side, head on the pillow, elbow on the mattress, about 35 cm from her face, slightly from above | The frame drifts and dips as her wrist tires and she shifts on the pillow | Low |
| Car (parked) | Driver seat, elbow resting on the door armrest, about 40 cm away, slightly from below. Headrest and side window behind her | The frame sways as she shifts in the seat | Low |
| Kitchen / desk / office | Standing or sitting at the counter or desk, elbow resting on it, about 40 cm from her face, slightly from below. Cupboards, tiles or shelves behind her | The frame sways as she shifts her weight and leans on the counter | Low-medium |
| Bathroom | Leaning on the sink, elbow on the edge, about 40 cm away, slightly from below; tiles and towels behind her, never the mirror | Small sway as she shifts her weight | Low |
| Street / park / walking | Arm bent, about 40 cm in front of her face at eye level, street or path and passers-by behind | The frame bobs with each step and rolls slightly side to side, exposure shifts | Medium-high |
| Bar / restaurant | Elbow resting on the table, about 40 cm away, slightly from below, people and lights behind | Moderate sway as she leans in and turns to friends | Medium |
| Gym | Sitting on a bench between sets, elbow on her knee, bent arm, people training behind | The frame moves with her breathing | Medium |
| Club / party / concert / festival | Arm bent close to her face, about 30 cm away, in the middle of the dancing crowd, lights behind | The frame moves with her bouncing to the beat, jolts when people bump her, drifts off her face and re-centers | High |

These rows are examples, not rules. Always adapt them with the real-filming brief (time of day, situation, energy of the script). Anything else → reason it out with the brief: the most natural handheld selfie for that place, the arm resting on something when people would, and a motion level that matches how lively the place really is.

**Push-in (one per video, motivated).** On the discovery line (the brand name, or the key line), a body action brings the frame up close: "she lifts her elbow off her knee and the frame pushes in close to her face, eye level, tight close-up, like a confidence". It lasts the whole line (2–4 s). The next beat brings it back: "she settles the elbow back on her knee, the frame drops back to the low angle with a small wobble". Adapt the body action to the rig (lifts her elbow off the armrest or the counter; while walking, "the frame pushes in close to his face"). Once per video, never more.

**Opening in media res:** the first beat starts with her already settled in the rig, already talking ("she rests there, a bit blasé"). No lead-in, no settling in.

Never write "part of her arm visible at the frame edge": models render it as a blurred giant forearm (seen in the Dutch test).

**Cuts between parts (multi-clip and extended videos).** Same framing on both sides of a join reads as one video glitching. Real creators cut and reframe, so every new part (Seedance extend or chained FLUX clip) opens with a jump cut to a slightly different framing: a bit closer or a bit wider, shifted a few degrees to one side, a slightly different height or tilt. Same person, outfit, place, light and voice; only the framing changes. Alternate the change from part to part (closer, then wider; left, then right).

### b. Beats: action, gestures and emotion per line

Split the script into **3–5 timestamped beats** (one sentence, or two short sentences that belong together, per beat). Each beat says, in this order: what her body does, what the frame does (if it changes), her emotion, then the line in quotes.

- **Emotion per beat** from the script: blasé or tired on the problem, a confidence on the discovery, brightening on the benefits, warm and sincere on the regret line and CTA. Write it as a plain word in the beat ("a bit blasé", "like a confidence", "brightens", "smiles").
- **Gestures: few, small, real.** 1–2 per 20 s from the free hand: "gestures with her empty right hand" on the benefits beat, "points down" on the CTA. A small everyday action on the last beat ("tucks a strand of hair behind her ear, smiles, then points down").
- Say which arm is out of frame and which hand is free ("her left arm is out of frame", "her empty right hand").
- Parts generated from a start frame or as an extension: one small gesture at most (more motion there distorts faces and hands).
- **No props in the hands.** The free hand is empty in every scene; props stay on surfaces (a mug on the counter, a coffee in the cup holder).
- **Never script eye movements** (no darts, glances, looks away, eyes widening): every scripted eye move came out distorted (Paris street test, Nov 2026). Never script eyebrows or eye shapes. Never an audible laugh.

### c. Voice: one line after the beats

"She speaks in [language] with a natural [accent] accent, [pace], like a FaceTime to a friend." Pace: "fast conversational pace" at home and in the car; late at night: "soft, low, calm voice, not a whisper"; loud places (club, concert, busy bar, street): "shouting over the [music / noise], voice straining a little". The voice identity never changes between parts.

### d. Actor and styling: one sentence

Nationality and age ("A French woman around 28"), hair with real texture ("messy shoulder-length light-brown hair"), makeup level that fits the scene (no makeup at home, light smudged makeup at a club), and real-life clothes for that place at that time (oversized hoodie and grey sweatpants at home, jacket in the car, gym top at the gym, going-out outfit at the club). Ordinary real person, never a model. Reused word for word in every part.

### e. Scene and light: specific and real

Name the place precisely with **3–5 concrete objects** that make it real and local (e.g. "a bright Parisian Haussmann bedroom: tall window with wrought-iron balcony, hanging plant, bedside table with a small lamp, cast-iron radiator"). Match the country of the script; never a named landmark. Busy places: people close around and behind her. Never a mirror or a phone among the objects.

One real light source in one short sentence:
- Daytime home: "Soft natural daylight from the window."
- Night or dim: "Warm light from the bedside lamp and a faint cool-white glow on her face." Never write "screen".
- Club: "Colored lights and strobes sweeping across her face."

### f. Audio: one line

- Quiet places: "Audio: only her voice recorded by the phone mic, close and dry, light room tone[, at most one faint real sound of that place and hour, e.g. distant traffic through the window], no music."
- Loud places: "Audio: raw phone mic overwhelmed by the [club / street]: [music pumping, bass distorting the mic / traffic], crowd right next to her, her voice slightly distorted but understandable."

Same audio line in every part.

## 3. Actor reference image ("image for my approval" and "use my image" only)

Skip this step entirely when the actor is generated natively. Before calling it, state the image cost in one line (step 0b). Use `arcads_generate_image_gpt` with `model: "gpt-image-2-5-sunburst"`, `nbGenerations: 2`, aspect ratio matching the video, and `referenceImages` = the user's actor and/or location images when given. Make an actor-in-scene selfie frame, never an ID photo:

> Frame from an iPhone front-camera selfie video. [Actor and styling sentence from 2d, or "the same person as the reference image, identical face"], [pose from the rig in 2a, e.g. "sitting on an unmade bed, knees pulled up, elbow resting on her raised knee"], mouth open mid-word. [Scene from 2e, or "in the same place as the location reference image"]. [Light from 2e]. Seen from her own selfie point of view, [distance and angle from 2a, e.g. "about 40 cm from her face, slightly from below"]; framing [framing from 2a, e.g. "head and shoulders, the ceiling and window visible behind her"], slightly off-center. One arm is out of frame; her other hand is empty. Eyes on the lens. Real unretouched skin, pores, stray hairs. Front-camera look: noise, soft blown highlights, front-camera softness, no cinematic shallow depth, no studio lighting, no beauty filter, no text.

Look at both takes with `arcads_view_frame` and keep the one that looks most like a real selfie frame (face matches the user's photo if given, styling and rig right, hands OK, no phone anywhere, no artifacts). Show the chosen image to the user and get a yes before generating video. If neither works, say why, state the cost of a new pair and ask before regenerating.

If the user's own image already shows the actor in the right scene and framing (and no phone), it can be used directly without a Sunburst step.

## 4. Duration, clips and beat timestamps: tight to the script

**The duration is the most important number in the whole skill.** It is computed from the script, line by line, before anything else is written, and shown in the plan. Never guess it, never pick a round number, never use the model's maximum.

Too long and the model invents: when the clip is longer than the speech, it fills the gap with fillers ("euh"), made-up sentences, a repeated last word, or a stretched last word, and may drop real lines (tested in French with the old block-style prompt: a 15 s clip for ~14 s of speech gave "simulez, simulez" on one take and "gratuit" stretched to "gratuiti" on another). Too short and the ending is cut off (tested with timestamped beats: 18.9 s of speech in a 19 s clip cut the last words twice; English, 7.9 s of speech in a 9 s clip cut the last word). The ending rule below sits between the two, leaning towards room for the last word.

- Estimate speech time:
  - English: ~2.8 words/sec.
  - French: count syllables, ~3.6 syllables/sec including short breaths between sentences (≈ 2.7 words/sec). Tested: 47 syllables → 14 s clip, clean. 51 syllables → 15 s clip, last word repeated or stretched (too long). Benchmark: 73 syllables → 20 s clip with timestamped beats and "fast conversational pace", word-perfect. 68 syllables → 19 s clip with timestamped beats, ending cut (too short); 68 syllables → 20 s, clean.
  - Dutch: count syllables, ~3.6 syllables/sec including breaths. Tested: 37 syllables → 10 s FLUX clip, word-perfect.
  - German / other: ~2.5 words/sec.
- **Count line by line:** for every sentence, write its syllable count (French, Dutch) or word count (English, German, other), its time, and the running total. Show this table in the plan. Recount once before generating; a counting mistake of 4 syllables is already 1 s.
- Adjust the rate to the delivery decided in the brief: fast conversational (home, car) = base rate; calm, tired or late at night = ~5% slower; shouting in loud places = ~10% slower.
- **Clip duration = speech estimate + 1.5 s for the ending, then always round UP** to the next whole second (cap at the model maximum). The beats' body actions (push-in, settling back, a stretch) slow the delivery, and the last word needs a moment to land. Tested Nov 2026 with timestamped beats: English, 22 words → 7.9 s estimate + 0.5 → 9 s clip cut the last word (car man); the rule now gives 7.9 + 1.5 = 9.4 → 10 s. 68 syllables → 18.9 s estimate → 19 s clip cut the last words twice (bedroom woman, car man); the rule now gives 20.4 → 20 s (cap). Benchmark: 73 syllables → 20 s (cap), clean.
- Never add more than that 1.5 s (plus the round-up), never a lead-in, never round up to a model maximum otherwise. Loud places: the ~10% slower rate already covers shouting, no extra padding.
- **Beat timestamps:** estimate each beat's speech time the same way, round to whole seconds, and lay the beats end to end from 0 s; the last beat ends exactly on the clip duration (give the extra time from the 1.5 s + round-up to the last beat, so the last line always has at least 1.5 s more than its own speech estimate). Format: "- 0-5s: … - 5-8s: …".
- **Never set the duration to the user's "max" by default.** A "20 s max" request with a 13 s script is a 13 s clip.
- If a take repeats, stretches or adds words at the END, the clip was too long: cut 1 s on the next take. If the END is cut off, the clip was too short: add 1 s.
- **FLUX, English** (native, ≤ 20 s): one clip only, 5–20 s.
- **Non-English**: FLUX clip 1 takes as much of the script as fits in ≤ 20 s (5–20 s, cut at a sentence end, duration from the rule above). The rest is split **only at sentence ends** into more FLUX clips of ≤ 20 s each, each started from the previous clip's last frame (e.g. 24 s of speech → two FLUX clips of ~12 s). Prefer evenly sized clips.
- **Seedance (English only)**: one clip up to 30 s (4–30). If the speech runs over 30 s, clip 1 takes as much as fits in ≤ 30 s (cut at a sentence end) and the rest is split **only at sentence ends** into Seedance extends of 4–30 s each, each with its own `duration` from the rule above. Prefer evenly sized parts.

Write the plan in one short message: model + reason, clip table (clip #, exact lines, duration with the line-by-line word/syllable count behind it, **estimated credits**), the **total estimated credits** (step 0b), and the creative direction from the real-filming brief (situation, rig + cause of motion, beats with emotion and push-in, actor and styling, scene and light, voice, audio, framing change at each cut). Ask "Go?" and generate only after the yes.

## 5. Generate

### FLUX path (`arcads_generate_video_flux_3_max`)

`resolution: "1080p"`, `generateAudio: true`, `reasoningLevel: "low"`, `nbGenerations: 1`, `aspectRatio` = chosen format (3:4 for 4:5), `duration` from step 4, clear `title` (e.g. `ugc_[brand]_[actor]_[location]_flux_v1`).
- `startFrame`: none for a native clip 1; the approved Sunburst or user image for clip 1 on rows N3–N4.

### Chained FLUX clips, non-English parts after 20 s (`arcads_extract_frame` + `arcads_generate_video_flux_3_max`)

For each part after clip 1: `arcads_extract_frame` on the previous clip with `frame: "last"`, then `arcads_generate_video_flux_3_max` with the same settings as clip 1 (`resolution: "1080p"`, `generateAudio: true`, `reasoningLevel: "low"`, same `aspectRatio`), `startFrame` = that extracted frame, `duration` from step 4, clear `title` (e.g. `ugc_[brand]_[actor]_[location]_flux_clip2_v1`). Launch the next part from the new clip right away. When all clips are done, `arcads_stitch_videos` [clip 1, clip 2, …] (0 credits).

Prompt: the extension version of the template (below), with "The person in the start frame" wording.

### Seedance path, English only (`arcads_generate_video_seedance_25`)

`resolution: "1080p"`, `audioEnabled: true`, `nbGenerations: 1`, `aspectRatio` = 9:16, 16:9 or 4:3, `duration` from step 4, clear `title` (e.g. `ugc_[brand]_[actor]_[location]_seedance_clip1_v1`).
- `referenceImages`: the actor image from step 3 (rows 4–5). Native actor: no `referenceImages`.
- **Motion reference (optional, strongest realism lever):** if the user supplies a real phone-shot UGC clip of the same setup (couch, car, street, club…), pass it in `referenceVideos` on clip 1 and add to the prompt: "Copy only the framing movement, handheld sway and body language of the motion reference video; not its person, place, outfit, words or audio." Use only clips the user owns or has the right to use. FLUX can't take it. For a seated bedroom/couch setup, the benchmark clip is already on Arcads: asset `6b87d179-5ba4-4b39-a887-b34d83b2ec6e`.

### Seedance 2.5 extend, English parts after 30 s (`arcads_extend_video`)

`model: "seedance_25"`, `videoPath` = the latest video (clip 1, then the previous extend's result), `duration` = this part's duration from step 4 (4–30), `resolution: "1080p"`, `audioEnabled: true`, `referenceImages` = the same actor image as clip 1 (rows 4–5; none for a native actor), clear `title` (e.g. `ugc_[brand]_[actor]_[location]_seedance_ext1_v1`). The result is the whole video so far, already stitched; launch the next extend on it right away.

Prompt: the extension version of the template (below).

### Prompt template (all paths)

Write it as short natural prose, like the benchmark, ~300–400 words, filled from the real-filming brief. Keep every sentence about actor, styling, scene, light, rig, voice and audio word for word across parts; only the beats and the framing change at a cut differ.

```
Authentic [vertical / horizontal / square] UGC selfie video filmed on an iPhone front camera, one continuous take, no cuts. [Actor and styling from 2d], [pose from the rig in 2a], in [scene from 2e with 3–5 concrete objects]. [Light from 2e]. Filmed from her own selfie point of view, about [distance] cm from her face, [angle], [what her elbow rests on]. Her [left/right] arm is out of frame. The only hand that ever enters the frame is her empty [right/left] hand. Framing: [framing from 2a, with what is visible behind her]. [Cause of motion from 2a, written as the frame moving, e.g. "The frame rocks slightly as her knee moves"]: organic micro-shake, small tilts, gentle sway.
- 0-[t1]s: [body + emotion, already talking]: "[line(s)]"
- [t1]-[t2]s: [push-in: body action + "the frame pushes in close to her face, tight close-up, like a confidence"]: "[discovery line]"
- [t2]-[t3]s: [the frame drops back with a small wobble; empty-hand gesture; emotion]: "[benefit line]"
- [t3]-[duration]s: [small everyday action, smiles, then points down]: "[regret line + CTA]"

She speaks in [language] with a natural [accent] accent, [pace from 2c], like a FaceTime to a friend. She says only these lines, word for word. Audio: [audio line from 2f]. Realistic skin texture, front-camera softness, auto-exposure shifts as the frame moves.
```

ACTOR wording by source (replaces the actor sentence):
- Native (FLUX clip 1, Seedance clip 1): the full sentence from 2d.
- Seedance with an actor image: "The person in the first reference image, same face, same hair, same outfit, …"; the scene ends "same place as the first reference image".
- FLUX with a start frame (rows N3–N4): "The person in the start frame, same face, same hair, same outfit, …"; the scene ends "same place as the start frame".

**Extension version (Seedance 2.5 extend or chained FLUX clip):** open with "Jump cut: the framing is now [slightly closer / slightly wider, a few degrees to the left / right]. The same person as in the source video (or the start frame for FLUX), same face, same hair, same outfit, same place, same light." Then the same point-of-view rig sentence, the beats for this part only, timed from 0 s, the first one already mid-conversation, one small gesture at most, no push-in. Then the same voice, "She says only these lines, word for word.", audio and realism lines.

Seedance with a motion reference video: add its copy line (step 5) after the rig sentence.

## 6. Run every part automatically, then hand over the final video

**After the user's "Go?" on the plan, run the whole pipeline without stopping, in the same turn**: clip 1, then every extend or chained FLUX clip (extract last frame → next clip) in order, then the final stitch for FLUX, polling `arcads_get_asset` (waitSeconds 40, repeat) between steps. FLUX takes ~2–8 min, Seedance ~5–8 min. **Never hand off to a scheduled task, reminder or `send_later` to wait**: keep polling until the final video exists. Never show intermediate clips and never ask for an OK between parts: the user only receives the final video.

**The user is the only reviewer.** Never analyze, transcribe, rate or QC a generated clip yourself: no `arcads_analyze_media`, no contact sheets, no probes, no frame checks on the output.

- User reports an issue on the final video: give the likely cause (duration too long for a repeated, stretched or added last word; too short for a cut-off ending; raw brand name or digit; punctuation; too much motion; scripted eye movements; echo or background noise for garbled words; a phone in the shot because the prompt gave it a hand, a position or an action, or made it propped, mounted or a mirror selfie), propose the fix, state the regeneration cost and ask "Go?". Never regenerate without the yes.
- If a generation fails (status `failed`): stop, state the cost of a retry and ask; on the yes, regenerate it once with the same prompt, bump `_vN`, and resume the pipeline automatically. Same failure twice: see the provider-block note in step 0b.

## 7. Joining the parts

- English, Seedance 2.5 extend: no trim, no stitch; each extend's result is already the whole video (tested: 11 s source + 13 s → 24 s output).
- Non-English, chained FLUX clips: each clip is separate, so stitch them all in order at the end (`arcads_stitch_videos`).
- Single clip: that clip is the final video.

## 8. Deliver

- Exactly one final video. Show it with the Arcads widget or the download URL from `arcads_get_asset` (signed links expire after 12 h: say so).
- One-line summary: models, duration, number of parts, resolution (1080p), and **actual credits used** (sum of `creditsCharged`) vs the estimate.
- Offer (don't do unasked): subtitles, music/SFX, translations, hook variants.

## Rules

- Scope: talking actor only, never a product. Product requests → point to a product skill in one line.
- **A phone is never shown.** Always a handheld front-camera selfie; never a propped, mounted or tripod phone, never a mirror selfie, never a second phone. **Phone wording:** "phone" and "camera" only in the first sentence and "phone mic"; the rig is a point of view ("Filmed from her own selfie point of view…, her left arm is out of frame, the only hand that ever enters the frame is her empty right hand"); every movement is the frame moving; never a line that names the phone to forbid it. Never write "phone screen" in prompts.
- Intake: one question per message, always. The format question is never skipped or assumed.
- Always show the estimated credits for the whole video in the plan and wait for one yes; then run everything automatically and report actual credits at delivery. Never regenerate without a yes.
- **Non-English script → FLUX 3 Max for clip 1 (≤ 20 s)**, any actor source; never Seedance or Kling. Beyond 20 s: more FLUX 3 Max clips of ≤ 20 s each, cut at sentence ends, each with the previous clip's last frame (`arcads_extract_frame`, `frame: "last"`) as `startFrame`; stitch all clips at the end. A specific actor → clip 1 start frame.
- **English script:** native actor → FLUX 3 Max if ≤ 20 s, Seedance 2.5 if 20–30 s; image for approval or user image → Seedance 2.5 with the actor image. Beyond 30 s: Seedance 2.5 extends (`arcads_extend_video`, `seedance_25`) of the latest video, same reference images; each result is already stitched, no trim, no stitch.
- "Generate natively" is ALWAYS the first and only recommended actor option.
- One model per path: English never mixes FLUX and Seedance (FLUX only for single clips ≤ 20 s); non-English is FLUX only, chained from the last frame.
- **Real-filming brief first, every video:** reason out how a real person would film this exact place, hour and script (situation, position, hold, angle, motion, gestures, light, voice, sound, cuts, local details) and remove the AI tells. The location table is a starting point, never a template to copy.
- **Duration computed line by line** from syllables or words, shown in the plan and recounted before generating.
- **Prompts follow the benchmark shape:** short natural prose (~300–400 words), specific place with 3–5 objects, point-of-view handheld rig with a physical cause of motion (Phone wording), then 3–5 timestamped beats (body + frame + emotion + line), then voice, "She says only these lines, word for word.", audio and realism lines. No walls of negatives.
- **Rig:** the filming arm rests on something when people would (knee, armrest, counter, table), about 40 cm, slightly from below at home and in the car, arm relaxed and never stretched out. Motion words always follow their physical cause.
- One motivated push-in per video, on the discovery line, held for the whole line, then the frame drops back with a small wobble. None in extension parts.
- Every new part (Seedance extend or chained FLUX clip) opens with a jump cut to a slightly different framing; same person, place, light and voice.
- **Duration is tight to the script and never set to the user's max by default:** speech estimate + 1.5 s for the ending, always rounded up, with all the extra time on the last beat. Beat timestamps add up exactly to the clip duration.
- Exact script only: never change the wording of a user's script beyond phonetic spelling and punctuation cleanup (proposed, not imposed), never add fillers. Never prompt for uneven pacing, hesitations, run-together words or laughs: the model reads them as permission to improvise.
- Brand names, URLs and numbers written out as spoken in the script language; no colons, ellipses or dashes in the dialogue; no pronunciation hints for words the model already says right.
- No named landmarks or real place names, no sweat or exertion wording: they can make the provider silently drop the output.
- Sunburst images are actor-in-scene selfie frames using the same point-of-view rig, with no phone visible, never ID photos, and are always approved by the user before video.
- Be in the action: the actor does what people really do in that place, never moved to a quiet spot. Motion, voice projection and audio match the real energy of the place.
- Real human behavior: 1–2 small empty-hand gestures per 20 s (one at most in start-frame or extension parts), one small everyday action, a smile on the last beat; no scripted eye movements, no eyebrows, no audible laugh. Never "arm visible at the frame edge".
- No props in the hands; props stay on surfaces.
- Camera rig, motion, beats, actor description, styling, scene, light, voice, audio and model are decided by the skill, not asked.
- No text, no subtitles and no added sound effects inside the generation (real location sound like club music is part of the scene). No post-processing pass. Always 1080p.
- **Never analyze or rate a generated clip yourself** (no analyze_media, contact sheet, probe or frame check on outputs). After the plan's "Go?", run clip 1, every extend or chained clip and the final stitch automatically in the same turn (never via a scheduled task) and deliver only the final video; no intermediate clips, no OK between parts.
- One final video, never multiple drafts.
