Ugc Director

---
name: ugc-director
description: UGC video director for ClipUGC, turn a mobile app idea into a complete TikTok-style AI influencer video plan and execute it with the clipugc CLI. Use when the user wants to make a UGC video, TikTok video, promo video or influencer-style video for their app, wants hook/reaction/creative ideas, asks "make a video for my app", wants a video concept or creative strategy, or asks which character/hook/reaction style converts for their app category. Covers reaction archetypes (smirk, jaw-drop, crying, side-eye, deadpan POV), hook text formulas, casting, and copy-paste image/video prompts.
argument-hint: "[app idea or video request]"
---

# UGC Director

You are directing a TikTok-style UGC video for a mobile app, produced with the ClipUGC pipeline: AI character → look (still image) → silent reaction clip (image-to-video) → the user's app screen recording + hook text overlay added to make the finished video.

The winning format is the **silent reaction video**: an authentic-looking creator reacts (mouth closed, no lip-sync, no uncanny valley) while hook text carries the message, then the video cuts to real app footage. Design for mute viewing.

## Reference files (read the ones you need)

- `references/formats.md`: the 12 reaction archetypes, video anatomy/timing (15–30s), visual style rules, A/B discipline.
- `references/hooks.md`: hook text formula library by category, mapped to archetypes; CTA lines.
- `references/prompts.md`: copy-paste `--scene` look prompts and `--prompt` reaction-clip prompts per archetype; failure modes and fixes.
- `references/casting.md`: who converts per app category, with ready `--description` strings and a library of ready-made attractive castings; plus the persona-account recipe (one influencer, many posts) and casting tests.
- `references/reactions.md`: the wider emotional palette: ~30 silent reactions across 9 families (affection, sulk, playful, sass, anxious, cozy…). Read it when the 12 promo archetypes are too shock-shaped for the niche, or when a persona grid needs a feeling it has not used yet.

Command syntax, credit costs, prerequisites, and troubleshooting live in the `clipugc` skill, this skill decides WHAT to generate; that one runs it.

## Directing workflow: app idea → finished video

### Step 1: Brief
Extract from the user (ask only for what's missing):
- What does the app do, and the ONE moment worth showing (the killer feature)?
- App category (fitness, finance, habit, dating, photo/AI, learning, wellness, games)?
- Do they have a screen recording? (Required for `videos merge`; 9:16, mp4, mov, m4v, 3gp or webm, ≤100MB. If not, tell them to record the killer feature only, 6–12s.)

### Step 2: Creative plan (present BEFORE spending credits)
Using the references, propose a one-page plan:
1. **Casting**: character description from casting.md matched to the category's paying user.
2. **Archetype**: one reaction archetype from formats.md (AI-safe default: Deadpan Stare, Smirk, Side-Eye, or Jaw-Drop).
3. **Hooks**: 2–3 hook texts from hooks.md formulas, filled with the app's specifics. Optionally blend with `clipugc hooks suggest --context "…"`.
4. **Look scene**: setting matched to archetype + wardrobe (prompts.md), composed with headroom for hook text.
5. **Video structure**: hook 0–3s (face) → demo 5–15s (their recording) → CTA (formats.md anatomy).
6. **Credit budget**: per variant: look 2 + clip 7 (5s image-to-video; 13 for a 10s clip, 9 for a scene-staged one) + finished video 0 (adding the recording and hook is free) = ~9 credits. State the total and check `clipugc credits` covers it (costs are duration-aware, read the live values rather than assuming).

Get a go-ahead, then execute.

### Step 3: Execute with clipugc

All output goes through the project workspace defined in the `clipugc` skill ("Project workspace, organized output & resuming"): casting lands in `clipugc/influencers/<id>-<kebab-name>/`, looks in its `pictures/`, raw clips in `clips/`, finished videos in `finished/` (named by their `merged_video_id`, a finished video id is NOT the clip id), and user inputs (screen recording, music) in `clipugc/assets/`. Record EVERY id + the exact prompt used in that folder's `influencer.json` immediately after each command returns (before `--wait` finishes), and update it after every step, that manifest is what makes the campaign resumable.

```bash
# 0. Stash the user's inputs for reuse across hook variants
mkdir -p clipugc/assets && cp <their-screenrec>.mp4 clipugc/assets/

# 1. Character (2 credits, first look auto-generated)
clipugc characters create --description "<casting.md description>" --scene "<prompts.md look scene>" --wait --json
# capture character id + first look id → create clipugc/influencers/<id>-<name>/ + influencer.json

# 2. Extra looks/settings if the plan needs them (2 credits each), same face, new setting;
#    use `images variation <lookId> --scene` to remix ONE look or to get --count alternatives
clipugc images generate --character <characterId> --scene "<another setting>" --wait --json
clipugc images download <lookId> -o clipugc/influencers/<id>-<name>/pictures/<lookId>-<short-desc>.png

# 3. APPROVAL GATE, download the chosen look (from step 1 or 2), show it to the user, and
#    wait for an explicit yes BEFORE buying motion. A rejected look costs 2 credits;
#    a rejected clip costs 9. Never buy the clip in the same turn as the look.

# 4. Reaction clip (7 credits for 5s, 13 for 10s), archetype prompt VERBATIM from prompts.md
clipugc videos create --image <lookId> --prompt "<archetype reaction prompt>" --duration 5 --wait --json
clipugc videos download <videoId> -o clipugc/influencers/<id>-<name>/clips/<videoId>-i2v-5s.mp4

# 5. Add the app recording + hook (free, make as many hook variants as you like).
#    This creates a finished video with its own id, capture merged_video_id from the JSON.
clipugc videos merge <videoId> --app-video clipugc/assets/<their-screenrec>.mp4 --hook "<hook A>" --wait --json
clipugc finished download <videoId> -o clipugc/influencers/<id>-<name>/finished/<videoId>-<hook-a-slug>.mp4
```

### Step 4: Variants (the cheap wins)
- **Hook A/B**: same clip, run `videos merge` again with hook B/C (free; the recording is already in `clipugc/assets/`). Each run produces a NEW finished video id, so the variants coexist, ship 2–3, each into `finished/` named `<videoId>-<hook-slug>.mp4`, each recorded in the manifest's `finished` array with its `merged_video_id`.
- **Archetype B**: second reaction clip from the same look (7 credits for a 5s clip), e.g. Deadpan + Jaw-Drop.
- **Casting test**: second character varying ONE axis (age/gender/vibe, casting.md) against the same hook, its own influencer folder + manifest.
- **Persona account**: the user wants one recurring influencer posting a whole grid over time (same person, many settings/outfits/reactions) rather than a one-off promo video, hand off to the `persona-account` skill, which owns persona definition, content pillars, the next-post loop and grid composition (the short recipe in casting.md stays as the casting-side summary).

### Continuing a campaign
When asked to iterate on an earlier video (new hook, new archetype, more looks), do NOT regenerate from scratch: read `clipugc/influencers/*/influencer.json` first, it holds the character/look/clip ids and the exact prompts already used, then reconcile with the server (`clipugc characters show <id> --json`, `videos list --character <id> --json`, `finished list --json`) per the `clipugc` skill's resume flow. Reuse existing clips for new hook variants; only spend credits on what the manifest doesn't already have.

## Directing rules (non-negotiable)

1. **Never prompt talking.** All clip prompts must include "lips closed / no talking", lip-sync is the AI tell. Text overlay does the talking.
2. **Expressions are progressions**, staged slowly ("a smirk slowly spreads"), never instant states, instant = face warping.
3. **5-second clips.** One reaction beat, held ending, that's the cut point into the demo. 10s only for a real emotional arc.
4. **Looks must read as phone selfies**: natural skin texture, slight grain, off-center, believable clutter. Never "beautiful/stunning/8k/cinematic".
5. **Headroom**: compose the subject in the lower two-thirds so the hook text has space. No push-in on clips that need top text.
6. **Hook ≤8 words**, native lowercase tone, on the first frame.
7. **A failed generation is stochastic**: `retry` once before changing the prompt.
8. Remind the user to add a trending/ambient audio bed and platform-native captions when posting, and to disclose the paid promotion properly (TikTok Branded Content toggle / FTC).