cinematic-trailer-pipeline

name: cinematic-trailer-pipeline description: End-to-end cinematic trailer production pipeline for Graeham Watts — from concept to finished MP4. Use ANY time the user mentions cinematic trailer, movie trailer ad, epic trailer, 60-second trailer, brand trailer, narrative video with VO, Drive-style video, Sicario-style trailer, multi-scene story video, "make me a trailer," "build a trailer," "trailer for my brand," "movie ad," "epic ad," or any request for a polished cinematic short-form video with narration, music, and a story arc. Also trigger when user wants to combine AI video generation (PAI / UTOPAI), ElevenLabs voiceover, and music into one finished piece. Encodes the full pipeline we built and the painful lessons learned — PAI content filter pitfalls, character drift workarounds, the editorial pass with color grade + music duck + VO mix, proper trailer pacing. Over-trigger — if the request smells like "epic cinematic video with a story," use this.

Cinematic Trailer Pipeline

The complete brand-trailer production system for Graeham Watts. Codifies the full workflow from concept to finished MP4. Built from the lessons of building "The Last 47 Days" — what worked, what failed twice, and the editorial craft that saved it.

When to use this skill

Use this any time the user wants a cinematic short-form video with narrative structure — typically a 30–90 second piece with:

• A story arc (setup → tension → climax → resolution)
• Multiple distinct scenes / shots
• Voiceover narration
• Music bed and sound design
• A title card / brand drop

Examples:

• "Make me a cinematic movie trailer ad. I want to be the star."
• "I want a Drive-style brand trailer for my real estate business."
• "Build me an epic 60-second video about [topic]."
• "Make a narrative video about [client situation] with my voice."

If the user wants a single talking-head video, route to heygen-video instead. If they want b-roll only, route to higgsfield-video. This skill is specifically for multi-scene narrative trailers.

The full pipeline (5 phases)

Phase 1: Concept Lock           → Tagline, dominant tone, 60s beat sheet
Phase 2: Character Refs         → Higgsfield Nano Banana Pro keepers
Phase 3: AI Video Generation    → PAI 2.0 (or fallback)
Phase 4: Voiceover              → ElevenLabs (narrator + Graeham clone)
Phase 5: Editorial Pass         → ffmpeg cut + color grade + music + mix

Each phase has its own section below. Don't skip phases — they build on each other.

Phase 1: Concept Lock

Before any generation, pin down four things. Use references/brief-template.md as the skeleton.

1. Dominant tone. Suspense-led usually wins for real estate. Action gets saturated fast (every realtor with AI tools made a Bond reel in 2025). Romance differentiates but underdelivers the "epic" feeling. The brief template defaults to suspense.

2. The anchor line. One memorable sentence that the whole trailer earns. For "The Last 47 Days" this was: "The story doesn't end the way it starts. It ends the way it's directed." The narrator delivers this at the resolution moment (usually 0:45–0:55).

3. The end card line. What the viewer takes away. For the 47 Days piece: "Selling your home is the highest-stakes story you'll ever live. I make sure it ends the way you wanted."

4. The 8-scene beat sheet. Write this out at minute-level granularity. Default structure for a 60-second trailer:

Adjust durations for shorter trailers (30s = compress each beat by half, drop one beat). For longer (90s+), expand the action and crisis sections.

Phase 2: Character Refs (Higgsfield Nano Banana Pro)

Why this matters: PAI 2.0's character module is only as good as the reference image you give it. A great selfie produces about a 25% face-match hit rate. Bootstrapping from a "keeper" first reference gets that to 75%+.

Workflow:

1. Start with a clean, front-facing, well-lit anchor photo of the user.
2. In Higgsfield, open Nano Banana Pro, upload anchor, generate 4 variants of the user in Look 1 (the dominant wardrobe — usually formal/overcoat for real estate trailers).
3. The user picks the keeper that "looks like them" (the user's eye is better than ours — proven twice this session).
4. Download the keeper as a JPG.
5. Bootstrap technique: Re-upload the keeper as the NEW seed photo. Generate 4 more variants at a TIGHTER framing (chest-up close-up). NBP locks much better when the seed is already in the target style.
6. Take the strongest 2 chest-up close-ups as the final ref material for PAI.

Reusable Higgsfield prompt template (Look 1 — formal arrival):

Cinematic photograph of the man from the reference image. Late 30s, neat short dark brown hair, light olive skin, clean-shaven. Wearing a [WARDROBE]. Standing in front of a [SETTING] at golden hour. Medium shot, three-quarter angle. Warm key light from camera left, soft fill, subtle anamorphic lens flare. Calm, watchful, deliberate expression. Cinematography reference: Drive (2011), warm shadows, slightly muted teal midtones. Shallow depth of field. Photorealistic, 35mm film look, subtle grain.

For Look 2 (different mood/wardrobe), keep the camera direction same, change wardrobe + setting + lighting style.

Output: 3 reference images (anchor selfie + Look 1 keeper + Look 1 close-up) in Downloads/<project>/refs/.

Phase 3: AI Video Generation (PAI 2.0)

PAI is the strongest narrative continuity tool, but it has pitfalls we hit hard tonight. See references/pai-gotchas.md for the full list. Top 3 you MUST know:

1. Content filter blocks emotional adjectives. Words like "tense," "suspenseful," "intense," "dread," "fear" trip a content policy filter that kills the render. Describe the visual mood instead: "low-key lit, quiet pacing, soft afternoon light, contemplative expression." Don't name the emotion.

2. PAI auto-overrides aspect ratio. When you mention "35mm anamorphic" in a prompt, PAI auto-switches to 16:9 even if the project is set to 9:16. Fix on the canvas node settings BEFORE running, or you waste 440 credits on the wrong aspect.

3. "Run" requires a manual click on the canvas node. PAI's chat will compose and ask for confirmation, then say "click Run on the canvas node." The Run button is the small up-arrow at the bottom-right of the composition panel. Auto-launch toggle is in the bottom bar of the chat — but enabling it requires accepting their IP rights modal, which is a per-action authorization the user must give.

Standard project setup:

• Aspect: 9:16 vertical for social trailers (most common)
• Quality: Pro 1080p (44 credits/sec → ~440 credits per 10s scene)
• Mode: Enhance (multi-shot decomposition) for narrative scenes; Raw for title cards
• Character module: upload all 3 ref images

Per-scene prompt structure:

• Lead with: "Scene: [setting]. [time of day]. [lighting]."
• Add: "Character [name] from the reference images, [wardrobe], [action]."
• Visual style: "Cinematic 35mm look, [color direction], Drive (2011) cinematography reference."
• Character lock: "CRITICAL: keep [name]'s facial features identical to the reference images."
• End with: "Generate a video clip, use Enhance mode." (or Raw for title cards)

Cost estimate: ~3,000 credits per 60s trailer assuming each scene runs first-try. Budget 50% more for retries. The Scene 1 / cold open is most likely to trip content filters — write it visually-focused first time.

Extraction: After all 8 scenes render, extract URLs via Chrome console JS:

[...new Set([...document.querySelectorAll('video')].map(v => v.src || v.currentSrc).filter(Boolean))]

Then bash download. See scripts/extract_and_download.py for the helper.

Phase 4: Voiceover (ElevenLabs)

Two voices, one rule: narrator voice for the cinematic frame, the user's clone for character dialogue. Don't try to use the clone for the entire narration — it sounds robotic because cloned voices are tuned for natural conversation, not trailer narration.

Standard voice selections:

• Narrator: Brian (nPczCjzI2devNBz1zQrb) — ElevenLabs stock, deep, measured, the closest thing to a film-trailer voice in the library. Alternatives: George (more British/gravelly), Adam (deeper/weightier). Let the user choose if they have a preference.
• Graeham's voice clone: Pa3vOYQHHpLJn1Tf7hnP — use for the character's actual spoken dialogue inside the story (e.g., the Scene 6 phone call line).

Voice settings — these are the ones that work:

• Narrator (calm/measured trailer voice): stability=0.40, similarity_boost=0.80, style=0.55, use_speaker_boost=true
• Graeham clone (natural dialogue): stability=0.35, similarity_boost=0.85, style=0.50, use_speaker_boost=true
• Do NOT use stability > 0.60 — that's what made the v1 clone sound robotic. Lower stability = more variation = more natural.

Use <break time="0.X" /> tags for cinematic pacing. ElevenLabs honors these on eleven_multilingual_v2. They give you deterministic pause control. Don't use <prosody> — silent pass-through.

Generation script: scripts/synthesize_vos.py — pass a JSON config with 5 lines, voice IDs, settings, output paths. Reads the API key from ~/Documents/Claude/Skills/.heygen-credentials/elevenlabs-key.txt.

Phase 5: Editorial Pass (the part that makes it land)

This is where the trailer goes from "AI footage with VO" to actually epic. The PAI clips alone are too long, the colors don't match between scenes, and there's no music. The editorial pass fixes all three.

What an editor actually does here

1. Re-time every shot. PAI gives you 5–10s shots. A trailer needs 2–4s shots. Trim to the meaningful moment in each, drop the setup tail.
2. Apply a unifying color grade. PAI's scene-by-scene color is inconsistent. One LUT across everything (warm shadows, slight teal midtones, light vignette) pulls the trailer into one visual world.
3. Music bed at ~30% volume. Music carries the emotion. Dial it under the VO so the words still cut through, but the score does the heavy lifting between lines.
4. VO mix at 1.4–1.7x. Narrator at 1.4x, character dialogue at 1.6–1.7x for punch.
5. Brick-wall limiter on master to catch any peak.

Standard cut sheet (60s source → ~40s trailer)

Adjust per project. The 40s target is good for social. For longer-form web video, keep more breathing room (50s).

Color grade presets (ffmpeg)

Three grades defined in references/ffmpeg-grades.md:

• Standard warm (Scenes 1–5, 7): colorbalance=rs=.06:bs=-.08,eq=contrast=1.10:saturation=0.92:gamma=0.96,vignette=angle=PI/5
• Cool crisis (Scene 6): colorbalance=rs=.04:bs=.02,eq=contrast=1.15:saturation=0.85:gamma=0.92,vignette=angle=PI/5
• Hero soft (Scene 7, alternative): colorbalance=rs=.05:bs=-.05,eq=contrast=1.08:saturation=0.95:gamma=0.97,vignette=angle=PI/5
• Title card: no grade (PAI nails text cards as-is)

Music sourcing

For a finished public deliverable, use these in order of preference:

1. CapCut library — if the user is doing the final edit in CapCut, search their cinematic music section. Strong titles: "Cinematic Trailer (Full Version)" by makesound, "Epic Drama Cinematic Trailer" by RomanSenykM, "An intense cinematic..." by Ryuichiro Yamamoto. License is bundled with CapCut export.
2. Pixabay Music — free, no attribution required, can download MP3 directly. Search "cinematic trailer" or "cinematic suspense." Best for when we're rendering the final externally (ffmpeg/Resolve).
3. Artlist or Musicbed — paid (~$30/track) but highest quality and intended for commercial cinematic use.

For ffmpeg-rendered trailers, default to Pixabay. The script in scripts/download_music.py has 3 known-good Pixabay cinematic trailer URLs as defaults.

The full editorial render

Use scripts/editorial_pass.sh. It takes:

• 8 scene MP4s (named scene1.mp4 through scene8.mp4)
• 5 VO MP3s (vo-01 through vo-05)
• 1 music track MP3
• A target output path

It produces a graded, cut, mixed, limited final MP4. The script is broken into two halves to beat sandbox timeout limits — render Part 1 (scenes 1-4) and Part 2 (scenes 5-8) separately, then concat + mix audio.

Honest reality checks (tell the user these upfront)

1. Character drift across scenes is unavoidable. Even with bootstrapped refs, you'll get 2–3 versions of the user's face across 8 scenes. The fix is editorial: fast cuts (under 4s each), lean on silhouettes and close-ups where the face is less recognizable, use a unifying color grade to mask differences.

2. Scene 5 (buyer / romance beat) is most likely to miss the brief. PAI tends to over-emphasize architectural shots (hardwood floors, marble counters) and underdeliver on the human moment. Plan to either regenerate Scene 5 with more aggressive human-focused prompting, or accept it and use editorial trimming to hide the worst part.

3. PAI's content filter will kill at least one scene per trailer. Budget for 1–2 retries on the alternate model. The cold open (Scene 1) is most vulnerable because it sets tension.

4. Music + VO + tight pacing = ~70% of the "epic" feeling. AI video alone, no matter how good, will not feel epic without these three elements. Do not deliver the trailer with just clips assembled — always do the editorial pass.

5. Total cost estimate:

   • PAI credits: ~$30–50 per trailer
   • ElevenLabs: pennies (we're well under the monthly char limit)
   • Music: $0 (Pixabay) or $30 (Artlist)
   • Time: 1.5–2 hours end-to-end if everything runs first-try; 3–4 hours with retries

Files in this skill

• SKILL.md — this file (you are here)
• scripts/synthesize_vos.py — ElevenLabs VO generation for the 5 standard trailer lines
• scripts/editorial_pass.sh — the ffmpeg cut + grade + mix pipeline
• scripts/download_music.py — known-good Pixabay cinematic trailer URLs + downloader
• scripts/extract_and_download.py — PAI clip URL extraction + download helper
• references/brief-template.md — the 8-scene trailer brief structure with prompts
• references/pai-gotchas.md — full list of PAI 2.0 pitfalls and fixes
• references/ffmpeg-grades.md — color grade presets + when to use each

Hand-off contract

Upstream: user request mentions "cinematic trailer," "epic trailer," "narrative video," or similar.

Downstream: finished MP4 in Downloads/<project-slug>/TRAILER-DIRECTORS-CUT.mp4 plus a folder of source assets (8 scene clips, 5 VOs, 1 music track, 3 character refs, brief.md).