blender-campaign-studio

Blender Campaign Studio

An agent skill that runs a commercial production as a pipeline with contracts: brief, shots, narration, music, right-to-left typography, Blender assembly, DaVinci Resolve finishing, QA. For Claude Code and Codex CLI.

Every paid generation is a ledger entry with a recorded approval and a cap. Nothing retries on its own. Hebrew is rendered, never typed into an NLE. A human watches and listens before anything is called done.

Built for Hebrew commercials. The typography path (Chromium bidi and shaping to PNG layers) should serve other right-to-left languages such as Arabic and Persian; only Hebrew has been verified.

Demo

The full 32-second film with sound: https://youtu.be/CB4LsyN5O_k

The film is a 32-second developer cut, assembled by the skill from its own plan (studio.py validate, studio.py assemble --render, delivery_qa.py). The Blender and DaVinci Resolve screens are real captures of the pipeline running, and the caption line below is what the typography path produces: a licensed Hebrew font shaped by Chromium into a PNG layer with the spoken word in amber, never typed into an NLE.

A 30-second Hebrew ad for the skill, docs/ad-30s-he-16x9.mp4, was made with the skill itself. One timeline JSON built the cut in Blender and was exported to DaVinci Resolve, where it was graded and rendered on the render station. The shots come from Grok Imagine, the narration and sound effects from ElevenLabs and the music from Suno, each through a recorded provider job, and every interface on screen is a real capture of that production. The mix measures -14 LUFS integrated.

Pipeline

flowchart LR
    A[Interview and brief] --> B[direction-v1.json<br/>shots-v1.json]
    B --> C{creative_gates.py<br/>direction / styleframes / paid-motion / rough-cut}
    C --> D[provider_jobs.py<br/>prepare - approve - execute once - collect]
    D --> D1[Grok via OpenClaw<br/>Genspark / Magnific / Dzine]
    D --> D2[HeyGen avatars]
    D --> D3[Codex stills]
    D --> D4[ElevenLabs v3<br/>word timestamps]
    D --> D5[Suno music]
    D4 --> E[narration_plan.py<br/>segments with pauses]
    E --> F[captions.py + render_text.py<br/>Chromium RTL to PNG]
    E --> G[audio_finish.py<br/>stems, ducking, envelopes, loudness]
    D5 --> G
    D1 --> H[plan.json<br/>Timeline JSON v1]
    D2 --> H
    F --> H
    G --> H
    H --> I[studio.py assemble<br/>Blender 5.1 VSE]
    H --> J[davinci_bridge.py<br/>FCP7 XML + verified import]
    J --> K[DaVinci Resolve Studio<br/>grade, Fairlight, render]
    I --> L[delivery_qa.py<br/>decode, loudness, hashes]
    K --> L
    L --> M[Master + viewing copy]

What runs locally, what needs an account

The core is local and free. Integrations add capabilities; each one is a route in connections.json that the skill probes before use and skips when absent.

Not included: the user-side clients the skill drives (suno-cli, heygen-cli, @genspark/cli, agy-cloudcode.mjs, the Magnific/Dzine clients, the OpenClaw runtime). Point connections.json at your own installations; it holds locations and model names, never credentials.

The ledger: how spending is controlled

python scripts/provider_jobs.py prepare --request shot-S04.json --jobs PROJECT/jobs        # semantic job id from the exact request
python scripts/provider_jobs.py approve-batch --jobs PROJECT/jobs --job ID1 --job ID2 ... `
    --evidence "user approved 8 shots, up to 4600 credits" --limit 4600 --unit "genspark credits" --per-job 900 --pilot 2
python scripts/provider_jobs.py execute --job PROJECT/jobs/ID1 --estimated-usage 460            # refused past the cap or outside the pilot
python scripts/provider_jobs.py batch-release --jobs PROJECT/jobs --batch BATCH --evidence "pilot reviewed: ..."
python scripts/provider_jobs.py collect-genspark --job PROJECT/jobs/ID1                         # downloads only what the receipt names

• A job is identified by the content of its request (prompt, settings, reference image bytes). Moving a file does not create a new attempt.
• execute runs once. A crash after submission leaves the job unknown; the next step is reconciliation against the provider's own receipt, never a resubmission.
• The batch ledger holds a lock while it checks the running total; a pilot holds the remaining jobs until a review is recorded.
• Style rules travel as positive style_constraints (up to 12) and at most three avoid items folded into every prompt of a series; antigravity.py check asks closed yes/no questions about each collected shot.

A timeline is data

plan.json (Timeline JSON v1) is what both studio.py assemble (Blender) and davinci_bridge.py export-xml (Resolve) consume:

{
  "version": 1, "width": 1920, "height": 1080, "fps": 24, "frames": 1152,
  "audio_target_lufs": -16, "audio_true_peak_max": -1.5,
  "clips": [
    {"id": "S01", "kind": "movie", "path": "shots/S01.mp4", "start": 1, "duration": 120, "channel": 1,
     "fade_in": 12, "motion": "zoom-in", "motion_amount": 0.05},
    {"id": "S01-freeze", "kind": "image", "path": "assets/S01-last-frame.png", "start": 121, "duration": 48, "channel": 1},
    {"id": "caption-1", "kind": "image", "path": "captions/phrase-001-state-01.png", "start": 20, "duration": 46, "channel": 10,
     "text_source": "captions/phrase-001-state-01.json"},
    {"id": "logo", "kind": "image-sequence", "path": "assets/logo-seq/frame_00001.png", "start": 1080, "duration": 72, "channel": 6,
     "hold_last": true, "resolve_movie": "assets/logo-resolve.mov", "fade_in": 6},
    {"id": "mix", "kind": "sound", "path": "audio-v001/master.wav", "start": 1, "duration": 1152, "channel": 20}
  ]
}

studio.py validate checks every file, trim, overlap, frame rate and envelope before anything renders, and warns when a small source is enlarged by a motion preset or when too many strips move.

Quick start

git clone https://github.com/Meir770ar/blender-campaign-studio.git "$env:USERPROFILE\.codex\skills\blender-campaign-studio"
New-Item -ItemType Junction -Path "$env:USERPROFILE\.claude\skills\blender-campaign-studio" -Target "$env:USERPROFILE\.codex\skills\blender-campaign-studio"
cd "$env:USERPROFILE\.codex\skills\blender-campaign-studio"
pip install -r requirements.txt
playwright install chromium
Copy-Item connections.example.json connections.json      # edit the routes you actually have
python scripts/studio.py doctor

A first production, end to end:

python scripts/studio.py init --project <productions>/my-film --idea "the idea in the user's words"
Copy-Item templates/direction-v1.json <productions>/my-film/planning/direction-v1.json     # replace every REPLACE: field
Copy-Item templates/shots-v1.json     <productions>/my-film/planning/shots-v1.json
python scripts/creative_gates.py --direction <productions>/my-film/planning/direction-v1.json --stage direction --out <productions>/my-film/planning/direction-gate-v001.json

Copy-Item templates/narration-v1.template.json <productions>/my-film/planning/narration-v1.json
python scripts/narration_plan.py split  --script <productions>/my-film/planning/narration-v1.json --out-dir <productions>/my-film/narration/requests-v001
# render request-single.json through provider_jobs.py (prepare / approve / execute), then:
python scripts/narration_plan.py layout --mode single --script <productions>/my-film/planning/narration-v1.json --jobs <productions>/my-film/jobs --out-dir <productions>/my-film/narration/layout-v001

python scripts/captions.py     --words <productions>/my-film/narration/layout-v001/words.json --style caption-style.json --out-dir <productions>/my-film/captions-v001 --fps 24
python scripts/audio_finish.py --spec <productions>/my-film/mix-v001.json --out-dir <productions>/my-film/audio-v001
python scripts/studio.py validate --plan <productions>/my-film/plan-v001.json
python scripts/studio.py assemble --plan <productions>/my-film/plan-v001.json --out-dir <productions>/my-film/render-v001 --render
python scripts/delivery_qa.py --video <productions>/my-film/render-v001/preview.mp4 --plan <productions>/my-film/plan-v001.json --audio-qa <productions>/my-film/audio-v001/qa.json --out <productions>/my-film/delivery-qa-v001.json

In Claude Code or Codex, a Hebrew request such as "בוא נכין סרטון לעסק שלי" activates the skill; SKILL.md is the doctrine the agent follows and references/long-form-commercial.md is a complete recipe for a 150-second commercial.

Status

Used daily by the author on Windows 11. 222 Python and 10 Node unit tests run in CI on Windows; they mock every provider and never render. Live capabilities of external services change: run the probe commands before a production and read the dated verification records in references/verification.md before trusting a documented limit.

Repository layout

Testing

python -m unittest discover -s scripts -p "test_*.py"
node --test scripts/test_grok_activation.mjs scripts/test_grok_native.mjs
ruff check scripts

scripts/verify_local.py exercises the real FFmpeg, Chromium and Blender paths on synthetic media.

Contributing, security, license

CONTRIBUTING.md for the development rules, SECURITY.md for reporting a vulnerability. MIT, see LICENSE; third-party code is listed in THIRD_PARTY_NOTICES.md.

תקציר בעברית

סקיל להפקת פרסומות וסרטוני תדמית בעברית מקצה לקצה: ראיון ובריף, בימוי וחוזי שוטים, קריינות עם השהיות מבוימות, מוזיקה וסאונד, כתוביות RTL בצבעי מותג, הרכבה ב-Blender 5.1, גימור ב-DaVinci Resolve Studio, בדיקות ומסירה. הליבה רצה מקומית בלי מנויים; כל ספק הוא מסלול אופציונלי. כל הוצאה בתשלום דורשת אישור רשום, אין ניסיונות חוזרים אוטומטיים, וכל שוט נצפה ונשמע לפני שהוא מוכרז גמור.