ugc-rockstar

Trigger

FIRES on: "GTA-style video", "make it feel like GTA V / Los Santos", "Rockstar aesthetic", "Vinewood crime short", "satirical fake-brand ad like the GTA radio", "heist-cinema look".

DO NOT FIRE when:

• The user points at one specific GTA-style video to reproduce → remix path (ralphy template use <slug>).
• The brief is a different aesthetic (analog horror, cinematic-narrative, clean UGC) → match that niche.

What this skill is

A generalized style overlay, not a finished video. It tells the scenarist and art-director HOW the Rockstar register reads so they produce a strong one for whatever premise the user brings. Runs through the normal pipeline and all gates.

Hard invariants

• Provider invariants stand. All generation via ralphy generate; renders via ralphy render <id>. No Higgsfield / Runway / raw API — those appear in the source draft and are NOT our stack. Read MODELS.md for the current i2v / image ids.
• Original homage, not impersonation. Build the aesthetic and satirical world — do not impersonate a real living person, and do not reproduce a specific copyrighted GTA mission/scene shot-for-shot. Fake brands and invented characters are the point.
• Music is a separate ElevenLabs pass with no artist/producer names (MEMORY.md) — describe the register (West-Coast hip-hop bed, cinematic tension score) by genre + instrumentation only. Ban music in the Kling prompt; overlay in the editor.

The GTA-V visual DNA (internalize before any output)

• Cinematic identity: late-2000s LA crime cinema (Mann / Scott / Fincher) filtered through satire. Hyperreal, slightly oversaturated, always slightly dangerous.
• Light: golden hour by default; even night is hot, neon-soaked.
• Camera: low angles for power, eye-level for vulnerability; slow-push establishing (wide, then creep in); rack focus on dialogue (bokeh'd luxury or grim urban texture behind a sharp face); the helicopter aerial pull-back (city as context for chaos).
• Grade: teal-and-orange pushed further — teal/cyan shadows, amber/gold highlights, skin popping orange.
• Tone — the double register: simultaneously funny AND real AND slightly dangerous. Never just funny, never just violent. A billboard, a radio line, or a fake brand in the background that comments on the world.
• Specificity: not "a luxury car" but "a Benefactor-style sedan in the hills, 7pm warm light"; not "a criminal" but "a man in a tracksuit outside an ammo store". The environment talks.
• Scene logic (the four levels): location → establishes character → creates conflict → reveals something about power/money, and through that, something universal. A chase isn't about the chase; it's about who owns the street.

Pipeline (enter at any stage)

1. Concept brief — title, logline, tone register (satirical ad / mission / short film / skit), characters, visual reference beats, runtime, platform.
2. Script — scene blocks: INT/EXT + invented location + time, action staging, dialogue in voice, one line of camera/lighting per scene.
3. Shot list — per shot: type (wide/medium/close/aerial/POV), subject, movement (static / dolly-in / pan / handheld / drone pull-back), duration, lighting/focus note.
4. Prompts — per shot, build the ralphy generate prompt: subject + action + invented environment + camera move + golden-hour light + the teal-orange grade as the style anchor. Fold the anti-AI-slop register in for any human shot (MEMORY.md).
5. Direction — grade, music register (separate ElevenLabs pass), pacing ("silence before violence" — build tension through quiet), editing rhythm.

Default model stack (verify against MODELS.md)

• Keyframes: google/gemini-3-pro-image-preview (premium typography for fake-brand billboards / on-screen text: openai/gpt-5.4-image-2).
• i2v: kwaivgi/kling-v3.0-pro for default cinematic motion; bytedance/seedance-2.0 for non-default physics (vehicle stunts, falls, chaotic action — per MEMORY.md VG model picks).
• VO / music: ElevenLabs; music a separate post-mix pass.

Workflow

1. Intake + announce: "GTA-V register — using the Rockstar skill." Pick the entry stage from what the user already has (concept / script / shots).
2. Produce the brief → script → shot list, getting a "go" at the storyboard lock.
3. Generate anchors one beat at a time with checkpoints, then i2v, then VO + music post-mix + captions, then ralphy render <id>.
4. Hand off to /evaluator.

Cookbook

• "Make a satirical GTA-style ad for a fake energy drink" → invent the brand + billboard, golden-hour hills, teal-orange grade, double-register VO. Match this skill, run the pipeline.
• "A Vinewood-at-night heist intro" → low-angle power shots, neon glow, silence-before-violence pacing.
• "Make it look exactly like that one GTA clip" with a link → remix path, not this skill.

Reference files (read on demand)

Deeper craft lives in references/ next to this skill — read the relevant one before producing that output type:

File	When to read
references/video-prompts.md	Building the per-shot ralphy generate prompts — style anchors, image + i2v templates, GTA location vocabulary, character archetypes.
references/cinematography.md	Shot patterns, camera moves, the teal-orange grade spec, editing rhythm.
references/characters.md	Character-voice profiles + visual archetype design.
references/satire.md	Fake brands, billboard copy, radio lines — the environmental layer.
references/techniques.md	Production techniques and continuity rules.
references/examples.md	Full worked examples — briefs, shot lists, prompts.
references/troubleshooting.md	Fixing output that doesn't read as GTA V.

All model/tool names in those files have been normalized to the Ralphy stack (ralphy generate, kling-v3.0-pro / seedance-2.0) — if you spot a stray foreign tool name, treat the Ralphy stack as authoritative.

See also

• docs/skills-vs-templates.md — niche-style skill vs remix template.
• MEMORY.md — anti-ai-slop, VG model picks (kling vs seedance), ElevenLabs no-artist-names.
• MODELS.md — current image / i2v model ids.