alecs5am/ralphy

Trigger

FIRES on a generic UGC-ad brief: "make a UGC ad for ", "creator reviewing my ", "testimonial ad", "problem/solution short for ", "before/after for " (when the emphasis is a talking creator, not a pure product montage).

DO NOT FIRE when:

• The user points at one specific ad to reproduce → remix path (ralphy template use <slug>). See docs/skills-vs-templates.md.
• The user wants to swap only the creator/person in an existing clip → /ugc-model-swap.
• The brief is a static image ad, not video → that is a different niche.

What this skill is

A generalized overlay, not a finished ad. It tells the pipeline HOW believable UGC ads work so the scenarist and art-director produce a strong one for whatever the user is selling. It runs through the normal pipeline and never bypasses the intake gates, the reference-required gate, or the quality gates.

Hard invariants

• Provider invariants stand. All generation via ralphy generate (OpenRouter media + ElevenLabs only — no FAL, no Higgsfield, no Canva, no raw API). Renders via ralphy render <id>. Read MODELS.md before naming any model id.
• Reference gate. A named real brand product ("iPhone 16", "CeraVe cleanser") fires the reference-required gate — refuse without a ref photo or logged --no-ref-consent. A no-name / generic product proceeds without a ref.
• AI-creator faces are synthetic personas, not real people. Build a consistent non-existent creator (see workflow). Do not face-mix toward a recognizable real person.
• Music is a separate ElevenLabs pass. Kling --audio is for the VO/lipsync only; ban music in the Kling prompt and overlay an instrumental bed in the editor stage (per MEMORY.md Kling-no-music note).

The niche, in one paragraph

A UGC ad lives or dies on believability and the first 3 seconds. The viewer must feel a real person is talking to them, not a brand. The hook is a problem mirror — show the viewer their own problem so they stop scrolling because they see themselves. Then a transformation/hope beat, then a real verdict, then a low-friction CTA. Trust comes from mannerisms (a nervous laugh, a hair tuck, direct eye contact, pointing at the product) far more than from polished delivery. The single biggest AI tell is a too-perfect, too-symmetrical creator and a studio-clean voice — fight both.

Beat / script shape (default ~15s, 9:16)

Author a shooting script as a table the art-director can fan out from:

Timestamp	Voiceover (exact words)	Visual / shot	Action / mannerism
0-3s	the hook line	problem shown visually, creator to camera	pattern-interrupt gesture
3-7s	the turn	product applied / used / demoed	hands on product
7-12s	the verdict	reaction / before-after	genuine reaction (laugh, raised brows)
12-15s	the CTA	creator to camera	point / "link in bio"

Virality principles to bake in: problem-mirror hook, hope loop (transformation after the problem), audio pattern-interrupt (a whoosh SFX on the cut), mannerisms = trust, room presence (a touch of reverb on the VO so it reads recorded-in-a-room, not studio), CTA urgency (specific, low-friction).

Framing + realism

• Handheld smartphone-UGC feel, eye-level, 9:16. Natural indoor light, not flat phone-flash.
• Photoreal creator. Skin pores, micro-detail, slight asymmetry — fold the anti-AI-slop / photoreal-still register in (ralphy guideline list; see MEMORY.md anti-ai-slop + photoreal-still notes). A flawless face is the giveaway.
• Generate the creator anchor with the product in hand or near the face for the first-frame reference.

Default model stack (verify against MODELS.md)

• Script: the scenarist LLM (callLLM) — feed it max context (product, target pains, the niche's viral patterns, the chosen hook framework).
• Creator + product keyframes: google/gemini-3-pro-image-preview (nano-banana-pro). Generate 3-5 variants, pick the most realistic.
• i2v (multi-shot talking creator): kwaivgi/kling-v3.0-pro — direct it shot-by-shot (each cut = creator speaks → product close-up → reaction → CTA). Kling --audio carries the EN VO + lipsync; for non-EN, VO via ElevenLabs (confirm target language at intake — see MEMORY.md Kling-no-RU-audio).
• Music: separate ElevenLabs Music pass, instrumental, post-mixed in the editor (no artist names — MEMORY.md).
• Captions: per-slot on the locked VO.

Workflow

1. Intake. Collect: product (name / URL / image), target audience + their pain, hook type (problem-solution / before-after / testimonial / transformation — or let the agent pick), target language, hard "no"s. Announce: "This is a UGC ad — using the UGC-ad skill." Defaults: 9:16, ~15s.
2. Reference gate. Named real brand product → get a ref or logged --no-ref-consent. No-name → proceed.
3. Script. Produce the timestamped shooting-script table; get the user's "go".
4. Creator persona anchor (with product), then per-beat anchors — one beat at a time with checkpoints.
5. i2v per beat (Kling, directed shot list), then VO/SFX, then music post-mix, then captions, then ralphy render <id>.
6. Hand off to /evaluator for the post-render quality gate.

Cookbook

• "Make a UGC ad for my no-name vitamin-C serum" → no ref needed; problem-mirror hook (dull skin), creator persona, ~15s 9:16. Match this skill, run the pipeline.
• "Creator ad for the Dyson Airwrap" → named real product → reference gate fires; ask for a ref photo first.
• "Same ad but with a different creator" → NOT this skill → /ugc-model-swap.

See also

• docs/skills-vs-templates.md — skill vs remix template.
• .agents/skills/intake/SKILL.md — the gates this runs through.
• MEMORY.md — anti-ai-slop, photoreal-still register, Kling no-music post-mix, Kling no-RU-audio.