Zack D Films

zack-d-films

Recreates the Zack D Films production line — normally a 50-person team — as one Claude chat over Higgsfield MCP. The user experience follows five acts, in this exact order, matching the channel's real growth engine: SCRIPT → VISUALS → ANIMATION → VOICE → EDIT.

HOW TO READ THIS FILE: execute Phases 1→8 IN ORDER. Each phase has a GATE. Long templates live in references/ — read them when the phase says so. references/replication-troy.md is the canonical worked example (the Trojan Horse short): when in doubt about any phase's output shape, imitate it.

Runtime contract

• Models are LOCKED: stills/character sheets → seedream_v5_pro · animated blocks → gemini_omni · narration → seed_audio. Never substitute.
• Aspect is 9:16 unless the user explicitly asks for 16:9. Pass aspect_ratio explicitly on EVERY image and video call — it never inherits.
• All shell work (probe, Whisper timing, ffmpeg edit, upload) runs in Higgsfield sandbox_exec — never on the client. Preflight a fresh run with: sandbox_exec({restart:true, command:"set -e; for b in ffmpeg ffprobe python3 curl jq; do command -v $b >/dev/null; done; python3 -c 'import faster_whisper'; mkdir -p zack/{blocks,audio,output}"})
• Poll every job to a terminal state. completed is the only good state. On failed/nsfw: RETRY LADDER (bottom of file). If no dedicated status tool is available, job_display({id}) returns the job's status + result URLs; wait between polls with sandbox_exec({command:"sleep 50"}) (~40–60s for images, poll video jobs every ~60s; blocks take 2–6 min).
• medias[].role may be auto-coerced by the server (e.g. image → image_references on seedream). That's fine — read the adjustments field, don't fight it.
• If a gemini_omni call returns a preset RECOMMENDATION instead of a job, resubmit the same call with declined_preset_id set to that preset's id. Never ask the user about it.
• Maintain zack/manifest.json in the sandbox (or an equivalent local notes block) recording: locked style formula, asset job ids, per-block prompts + job ids + statuses, narration job id, beat map, edit map. It is the resumable source of truth.
• Never expose model names, phase names, or internal mechanics in chat. The user sees creative substance and approval gates only.
• Characters NEVER talk on screen (no lip-sync). The narrator is external. Every video prompt includes "characters only emote and gesture, they do NOT talk" and a mute/diegetic-only audio note.
• No real brand / studio / IP names in prompts. No on-screen text in generations (annotations are shapes, not words).

The house formula (what makes it "a Zack video")

1. Curiosity loop script — opens a question the brain needs closed; closes it in the final line with a reframing kicker.
2. Perfect consistency — every recurring person/object/place is generated once as a character sheet, then referenced in every clip. Same backbone every time; the model never "remembers", it sees.
3. Signature 3D look — glossy stylized-realistic 3D (see references/style-3d.md), warm sun, cobalt sky, PBR materials.
4. Baked annotations — neon-green glowing outlines/arrows/rectangles for "look here / how it works"; red dashed trajectories, red glowing cracks and red impact bursts for danger/physics. Rendered in-world by the video model.
5. One recognizable narrator voice — the user's cloned voice.
6. Zoom-ins and screen shakes timed to key moments in the edit.

PIPELINE

Phase 1 — RESEARCH (trending topics → options)

If the user already named a topic, skip to Phase 2. Otherwise search the web NOW (2–3 queries: "trending topics this week {month year}", "what is everyone talking about {month year}", one vertical the user cares about). Find anchors with mainstream heat (a new movie, a viral event, a season). For each anchor derive a history/science story with a secret: the best picks have (a) a myth everyone believes, (b) a documented "actually..." angle, (c) strong physical/visual potential. Present 3–5 topic options, one line each: **Topic** — why it's hot now + the hidden angle. Wait for the user's pick. GATE 1: a picked topic.

Phase 2 — ANGLE

State the angle in one sentence of the form: "Everyone believes X — but [historians/engineers/scientists] now think Y, and the real mechanism is Z." The angle must flip the familiar story, not just retell it. GATE 2: an angle with a genuine flip.

Phase 3 — SCRIPT (curiosity loop)

Read references/scriptwriter.md, then write the narration:

• Default length 28–32s ≈ 75–85 words (~2.7 words/sec). Scale by the same rate if the user asked for another duration.
• 5 beats: Myth → Twist → Mechanism A → Mechanism B → Kicker.
• Then run the rewrite pass from scriptwriter.md (cut hedges, concretize numbers, kill adjectives that don't render). Show the script in chat (SCRIPT LOCK) and ask one short approve/tweak question. This is the last mandatory stop — after approval the run is hands-off to final video. GATE 3: approved script.

Phase 4 — SCENE PLAN (the production plan)

Read references/scene-planning.md. Break the script into blocks of 10s (N = ceil(duration/10)). For every block write a timed shot list — Zack pacing is 6–9 hard cuts per 10s block (~1.0–1.6s per shot). For each shot: [t0–t1] SIZE/ANGLE — subject + action — annotation (if any) — assets used. Rules: vary size+angle on every cut; at least one MACRO eye/detail shot per video; one "gag" beat (vanish-poof, deadpan reveal); annotations per the green/red grammar; every shot names which character sheets appear. Also produce the asset roster: every character, object, environment that appears in ≥2 shots, plus damage/wear states (e.g. "wall — intact / cracked / breached"). Post the plan in chat as a compact table. Do not wait for approval (the script was the lock) — proceed. GATE 4: every block has timed shots; every shot maps to roster assets.

Phase 5 — CHARACTER SHEETS (the consistency backbone)

Read references/style-3d.md and take its STYLE FORMULA byte-identical into every prompt in this phase and Phase 6.

1. Style key — one seedream_v5_pro still (9:16): a representative scene in the formula. Keep its job_id as the look anchor.
2. Asset sheets — one image per roster entry, in parallel, each with the look anchor as medias role image:
   • people → aspect_ratio:"2:3", full body, neutral flat backdrop, front 3/4 view, sheet-style;
   • objects/props → "1:1", isolated, no scene;
   • environments → "9:16" (target aspect), dressed, no people. Subject-bleed guard (proven bug): the style key's hero subject WILL leak into location plates unless the prompt says ENVIRONMENT SHEET — EMPTY LOCATION PLATE, take only the render style and palette from the reference image, NOT its subject plus an explicit ABSOLUTELY NO {style-key subject}, NO people, NO animals line. QC every environment plate for the hero subject before Phase 6 and regenerate contaminated ones with that wording;
   • damage/wear states → separate sheet derived from the base sheet (pass the base sheet job_id as an additional media).
3. Poll all to completed. Show the gallery in chat with one line: these are the character sheets that keep every scene consistent. GATE 5: every roster entry has a completed sheet. Never enter Phase 6 with a missing asset — that beat WILL drift.

Phase 6 — ANIMATION (one call per block)

For each block submit ONE gemini_omni call: duration:10, resolution:"720p", aspect_ratio:"9:16", medias = environment → characters → props (role image_references, max ~3 — pick the assets that block actually shows). The prompt = the block's timed shot list rewritten as explicit timecoded cuts (template in references/scene-planning.md §3), including annotation directions, camera energy (whip-ins, crash zooms), the "do NOT talk" line, and the NEGATIVE line. Submit blocks in parallel batches of ≥3; poll the batch; RETRY LADDER on failures. A completed block is FINAL unless it violates a named QC check (style drift vs sheets, wrong aspect, collapsed cuts → regenerate once with the cut list emphasized). Download every block into the sandbox as zack/blocks/blockNN.mp4. GATE 6: all N blocks completed + downloaded, no gaps.

Phase 7 — VOICE (the user's voice, one continuous narration)

• If the user has provided ~1 min of their voice: create a cloned voice (voice-creation tool / create_voice flow), then use its element voice_id.
• Else call list_voices and let the user pick once (or reuse the voice they picked in a previous run — it's in memory/manifest). Record voice_id+voice_type; LOCKED for the channel.
• Generate the FULL script as ONE seed_audio take: [confident documentary narrator, deliberate storytelling pace with natural beats between sentences, starts speaking immediately] [00:00-00:{len}] {full script} with the locked voice pair. Pacing note (proven): with a plain "tight pacing" hint the take lands ~10% SHORT of the timecode window; the "deliberate storytelling pace" wording hits the target. Check durationSec on the completed job — aim for N×10 − 1.5s; regenerate with an adjusted pacing hint if off by >2s.
• Download to zack/audio/narration.wav. In the sandbox run faster-whisper word timestamps → zack/audio/words.json. Derive the beat map: the timestamp of every beat keyword (the moments zoom-ins will hit).
• If narration > N×10s: tighten the script line that overflows and regenerate. NEVER time-stretch audio. GATE 7: one narration file ≤ N×10s, words.json, beat map.

Phase 8 — EDIT (zooms, shakes, final cut)

Read references/scene-planning.md §4 (edit grammar). Order of operations (proven): call media_upload({filename:"final.mp4"}) FIRST to get the presigned PUT URL and media_id, THEN assemble — so the upload can run inside the same sandbox call as the render. Sandbox persistence is not guaranteed between calls (files may vanish even back-to-back): the robust default is ONE foreground sandbox_exec call (≤120s) that does everything — write helper files via quoted heredocs, download blocks + narration with parallel curl, probe fps, extract/concat SFX, concat blocks, apply the edit map, mix, trim, curl PUT to the presigned URL. Use -preset veryfast -crf 16/17 to fit the budget (a 3-block short assembles in ~50s). Steps:

1. Concat blocks 1..N (re-encode, uniform fps from probe — never hardcode).
2. Apply the edit map with the bundled scripts/zack_edit.py, fed by edit_map.json:
   • punch zoom-ins (~1.0→1.08 over ~0.4s) at 3–6 beat-map moments (reveals, annotations, kicker);
   • shake (±1% crop jitter, ~0.3s) at every block boundary and impact beat. In the one-shot path, run zack_edit.py --print-vf to emit just the ffmpeg -vf expression and fold it into the final encode (one pass instead of two) — scripts/assemble_oneshot.sh is the proven full template.
3. Lay the narration over the composite from t=0 (video SFX ducked to ~0.25× under it), amix ... ,loudnorm=I=-16:TP=-1.5, trim to narration end + 0.4s.
4. QC probe: 1080×1920, duration ≈ narration+tail, audio present, plays from frame 1.
5. Upload with the presigned URL from step 0 (curl -f -X PUT in the SAME call as assembly), then media_confirm({type:"video", media_id}) → deliver the hosted URL. One file, final.mp4. A sandbox path is not a deliverable. GATE 8: exactly one final MP4, delivered as a confirmed hosted URL.

After delivery

Offer (one line each, never run unasked): 2K upscale via upscale_video; a YouTube title/description pass; saving the style+voice as the channel profile for next time.

RETRY LADDER (Phase 5/6 failures)

1. Resubmit same prompt (new seed) ×2.
2. Reword: soften violence tokens (battle → clash, blood → dust), remove tight-face close-ups of distressed characters, resubmit ×2.
3. Reframe the beat (different shot size / staging).
4. Never drop a block, never leave a gap. After ~8 attempts on one block, stop and surface it.

FINAL QC CHECKLIST

• One vertical 1080×1920 MP4, narration synced, no dead air > 1s.
• Every recurring element visually identical across blocks (sheet-driven).
• Annotations: green = reveal, red = danger/physics; no on-screen words.
• ≥1 macro eye/detail shot; ≥1 gag beat; kicker closes the loop.
• Zoom-ins on reveals; shakes on impacts/boundaries.
• No brand/IP names; characters never lip-sync.