One engine-agnostic system: story → cinema → AI generation → continuity → QA → edit → post → distribution → learning
An AI audiovisual production engine that replaces empty adjectives with real optical and color parameters — FOV in degrees, Kelvin, 180° shutter, T-stop, saturation 0–100, HEX tints — instead of ultra real, 8k, masterpiece.
62 skills · 69 Visual DNA styles (each with Color Grading DNA) · 25 pipelines · 22 profiles · 31 engines + 20 adapters · 35 task routes · 13 quality gates · 16 style blends · 21 LUT presets
Credits: Daniel Rodrigues · Draft Creative Studio Ltd
Runs on Claude, ChatGPT, Gemini, DeepSeek, Qwen, Kimi and any capable LLM — see UNIVERSAL_PROMPT.md.
Download the manual: https://github.com/iwerry/CineOS_3.2/blob/a651199d9b0c62300187c713e34c33b25e833d62/CineOS_3.2_Manual.pdf
Cinema OS is a skill that turns ordinary requests ("make a Reel about my coffee shop") into structured professional production: script, storyboard, shot list, per-engine generation prompts, synced audio, film-emulation post, editorial calendar and performance analysis.
A prompt is not guesswork. It is cinematography, written down.
The engine never describes a scene with pretty adjectives. It resolves the shot into geometry, physics and color science: field of view in degrees, motivated camera movement, one dominant light source, Kelvin temperature, contrast ratio, HEX palette, numeric saturation and tint, and diegetic audio with timestamps.
Visual identity is locked before generation (Style Bible) and composition is locked in a still before animation (Hero Frame First). That is what separates an AI-looking clip from a clip with a director's intent.
Not a bigger skill. A smarter cinematic system. See
ARCHITECTURE.md.
| Layer | What 3.2 adds |
|---|---|
| Cinematic Memory | Skill 65 Project Bible (PROJECT.md = truth vs STYLE_BIBLE.md = language, 16 templates in PROJECT_BIBLE_TEMPLATE/) · Skill 70 Continuity Graph + AssetGraph ("if Helena's hair changes, which shots regenerate?") · Skill 74 Learning Loop |
| Intent ≠ syntax | Skill 66 Shot DNA + Prompt Compiler: one ShotSpec → compileShot() → Veo / Kling / Seedance / Hailuo / Wan / Hunyuan / Higgsfield / Flux / Midjourney / ComfyUI dialects through 20 engine_adapters |
| Model Intelligence | Skill 67: best model for this shot, now, with these assets and budget + a 10-test benchmark suite (T001–T010, results empty until you measure) |
| Direction | 59 Cinematic Grammar · 60 Cinematography Director ("why this lens?") · 61 Acting Director 2.0 (beats, subtext) · 62 Sound Cinema Engine (silence as an event) · 69 Edit Engine |
| Audit → Critique → Polish | 63 AI Artifact Detector · 64 Cinema Audit (12 weighted dimensions) + Cinema Slop Detector (15 rules) · gates G10–G13 · /cinema:polish regenerates only failing shots |
| Professional pipeline | 71 Software Workflow Bridge (Premiere, After Effects, Photoshop, Illustrator, DaVinci Resolve, CapCut, Blender, Higgsfield, Magnific, Topaz — TOOLKIT_2026.md) · 72 Deliverables Studio (screenplay, treatment, pitch deck, storyboard PDF) · 73 Code-Driven Animation · 75 Micro-Drama Showrunner |
| Look development | 10 new Visual DNAs (CINEMANOVO, SERTAOBLEACH, TROPICALNOIR, PRECISIONTHRILLER, SLOWCINEMA, WUXIAINK, DONGHUA3D, SOLARPUNK, BRUTALISTSCIFI, VERTICALDRAMA) · 4 blend recipes · 9 LUT presets · Director Profiles · Cinematic Design System (tokens) · Skill 68 Reference Intelligence |
| Engines | +8 (Magnific, Topaz Video AI, FLUX Kontext, Qwen-Image, HunyuanVideo, Ideogram, Recraft, Adobe Firefly) — confidence low, re-verify |
| People | Profiles 19 DoP · 20 Editors/Colorists/VFX · 21 AI Producers & Micro-Drama · 22 Screenwriters & Pitch |
New commands: /project:init /project:audit /story:develop /story:analyze /shot:create /shot:audit /scene:direct /camera:design /acting:direct /reference:analyze /continuity:audit /model:benchmark /model:recommend /prompt:compile /cinema:audit /cinema:critique /cinema:polish /cinema:finish /edit:cut /artifact:check /code:animate /deliver:screenplay.
import { compileShot, compileShotForAll, modelIntelligence, cinemaAudit, polishPlan, cinemaSlop, AssetGraph, sequenceContinuity } from './skillsData';
compileShot(shot, 'veo_3_1'); // intent → adapter → prompt + warnings
modelIntelligence('S12', { kind: 'video', dialogue: true, durationS: 8, budget: 'medium' });
cinemaAudit({ story: 91, continuity: 71, /* … */ }, findings); // score + CRITICAL/WARNING + polish list
new AssetGraph().add(...shots).impactOf('Helena');Honest status. Benchmark results and the learning loop are schemas until you feed them real data. Engines added in 3.2 are from training knowledge (verified 2026-06-30, confidence low). Some v3.1 strings in the JSON and the v3.1 catalog tables remain in Portuguese. Run python scripts/validate_contract.py and tsc --strict after edits. Audit of the base: CINEOS_AUDIT_v3.1.md.
| Upgrade | What it gives you |
|---|---|
| 🎨 Color Grading DNA | Every one of the 59 styles now carries color_grading: saturation (0–100), contrast curve, shadow_tint and highlight_tint (HEX), grain structure, dynamic_range and the closest lut_match. Plus a tables.color_science reference (saturation scale, contrast vocabulary, grain glossary, tint rules). |
| ➕ 12 new Visual DNAs | UFOTABLE, MAPPA, WEBTOON, ISEKAI, GHIBLIPUNK, GLASSNOIR, CLAYMORPHISM, LIQUIDMETAL, GLITCHCORE, DREAMCORE, ANALOGDREAM, BIOLUMINESCENT — modern anime compositing, kinetic sakuga, vertical webtoon, isekai key-visual, and the 2025–26 AI-native looks (Flux-era analog, liquid metal, glitchcore, dreamcore, bioluminescent, claymorphism). |
| 🧪 Skill 57 — DNA Blending Lab | Blend exactly 1 base (60–70%) + 1 accent (30–40%), with the accent isolated on a single element. Optics inherited from the base, palette and grade interpolated, negative locks merged, conflicts flagged. blendStyles() in TypeScript. |
| 🎚️ Skill 58 — Color Grading DNA Engine | Turns a style into a numeric grading card wired into the Darkroom chain (Skill 25). gradeCard('UFOTABLE') returns color_grading_card.json matched to a LUT preset. |
| 👥 2 new profiles | perfil_17 Independent Animation, Anime & Webtoon · perfil_18 Art Direction & Signature Look. Existing profiles gained the new styles and skills where they fit. |
| 🔀 New pipeline & route | p_signature_style and blend_style — build a reusable, ownable look for a client or channel. |
| ✅ Smarter lint & veto | New GRADE_VAGUE (grade without numbers) and BLEND_OVERLOAD (3+ styles) checks; image analysis now returns color_grading_estimate; gate G1 now requires the grading card. |
| 🎞️ 2 new LUT presets | Portra 400 −0.5EV Analog Dream and Anime Compositing Neutral+FX. |
# Generic route — the engine routes by itself
/route "I want a 9:16 Reel of my coffee that never gets cold, 80s style"
# Explicit pipeline
/pipeline:p_social_reel --style SYNTHWAVE --ratio 9:16 --prompt "the coffee that never gets cold"
# NEW: blend two Visual DNAs into a signature look
/blend:NOIR+LIQUIDGLASS --weights 70/30
# NEW: get the numeric grading card for a style
/grade:UFOTABLE
# Reverse-engineer an image and get a recreate-prompt per engine
/route "analyze this image and give me the recreate prompt for Veo"
# Real brand or product in the scene? Brand Mode (opt-in)
/brand:on --marca "X" --produto "Y" --vinculo afiliadoSmart autoload: Skill 15 (foundation) always · 16 for ComfyUI · 25 when post is requested · 53 on every video · 58 when color/saturation/shadows/grade is mentioned · 57 when two styles appear · 55/56 before delivery · G9 whenever a real brand or product is cited.
Core five below; v3.2-CIA adds ARCHITECTURE.md, CINEOS_AUDIT_v3.1.md, TOOLKIT_2026.md, UNIVERSAL_PROMPT.md, PROJECT_BIBLE_TEMPLATE/, scripts/validate_contract.py, CHANGELOG.md.
| File | Role |
|---|---|
BaseSkill.md |
🏛️ The constitution: execution protocol, 16 universal laws, optical + color vocabulary, engine knowledge, task modes, quality gates |
draft_studio_config.json |
💾 Single source of data: models, profiles, skills, styles, pipelines, routes, resolved tables, templates, characters, clients |
skillsData.ts |
⚙️ Types + typed loader + resolution engine: resolveStyle, routeTask, recommendEngines, compilePrompt, lintPrompt, profileKit, blendStyles, gradeCard, helpers |
skills_cinema_pipeline.md |
📚 Catalog: 45 skills + 59 styles + Color Grading table + 18 pipelines + methodology |
profiles_guide.md |
👥 18 profiles with pains, technical DNA, skills, styles, engines, pipeline and KPIs pre-resolved |
ID contract (never break it): skill_NN (01–58; 32–44 = legacy styles), dna_<alias> + upper-case ALIAS, perfil_NN, p_<pipeline>, <engine> (e.g. seedance_2_5), gates G1–G9, routes by id. Every .md cites only IDs that exist in the JSON, and the TS reads the same JSON — the docs never lie about the data.
| Layer | What it is | Lives in |
|---|---|---|
| Skill | an executable ability (script, route an engine, grade, blend) | skills[] |
| Knowledge | lookup tables: optics, light, camera, engines, markets, color science | tables{}, models[] |
| Visual DNA | a reusable look kit: optics + palette + texture + light + motion + color grading | styles[] |
| Blend recipe | a 70/30 hybrid of two DNAs | tables.ai_style_mixes |
| Profile | who is asking + the default kit | profiles[] |
Rule of thumb: new look → style · new capability → skill · new problem owner → profile · new reference data → table.
1 PRE-PRODUCTION 47 Story → 21 Script → 52 Bible → 22/23 Character → 45 Reference analysis → 57 DNA blend → 48 Storyboard
2 PRODUCTION 46 Hero Frame → 53 Engine per shot → 19/20 Direction & camera → Visual DNA → 58 Grade → 24 Acting → 26/49 Audio
3 POST 54 V2V (ordinary video → cinema) → 25 Darkroom/LUT → 26 Mix → 55 QA
4 DISTRIBUTION 17 Hook → 18 Thumb → 29 Calendar (3 hashtags) → 31 Localization → 56 Compliance
5 ANALYTICS 30 Retention & CTR → back to 17/18/21/28
- Route the request →
routeTask() - Profile →
perfil_01–18inherits styles, engines and pipeline - Style →
resolveStyle()— max 1 base + 1 accent (blend via Skill 57) - Skill chain →
expandSkillChain()with dependencies - Style Bible (G1) → alias + HEX palette + optics + grading card locked
- Hero Frame First (G2) → approved still before animating
- Feasibility Veto (G3) → does the plan break physics? Rewrite the plan, not the prompt
- Engine per shot →
recommendEngines()withverified_on+status - Compile + Lint (G4) →
compilePrompt()→lintPrompt() - Deliver the artifact + one line of assumptions + next steps
Each style is a technical kit, not a pretty reference: FOV°, camera, lens, aperture, shutter, WB, fps, HEX palette, textures, motion language, light, prompt_core, color_grading and engine affinity per family.
| Family | Aliases |
|---|---|
| Comics | MIGNOLA · SINCITY · MOEBIUS · WEBTOON ⭐ |
| Anime | AKIRA · GHIBLI · SHINKAI · TRIGGER · LOFI · UFOTABLE ⭐ · MAPPA ⭐ · ISEKAI ⭐ · GHIBLIPUNK ⭐ |
| 3D animation | PIXAR · SPIDERVERSE · ARCANE |
| Classic / craft | XEROX · UKIYOE · PIXELART · LAIKA · CLAYMATION · PAPERCUT |
| Auteur cinema | KUBRICK · FINCHER · VILLENEUVE · WONGKARWAI · WES · NOIR · LEONE · BURTON · NEWHOLLYWOOD70S · GLASSNOIR ⭐ |
| Pop & retro | KPOP · SYNTHWAVE · Y2K · VAPORWAVE · SUPER8 · GLITCHCORE ⭐ |
| Documentary | DOCREAL · PLANETEARTH · TRUECRIME · ANALOGDREAM ⭐ · BIOLUMINESCENT ⭐ |
| Commercial | KEYNOTE · CHEFSTABLE · EDITORIAL · SPORTS · LIQUIDMETAL ⭐ |
| Design / UI / Edu | SWISS · ARTDECO · BLUEPRINT · NEUBRUTALIST · LIQUIDGLASS · CLAYMORPHISM ⭐ · KURZGESAGT · WHITEBOARD · STORYBOOK |
| Horror | FOUNDFOOTAGE · DREAMCORE ⭐ |
⭐ = new in v3.2
/style:MIGNOLA # by alias
styleAlias: 'SINCITY' # in TypeScript
/style:dna_wes # by id"color_grading": {
"saturation": 80,
"contrast": "very high, CG-light overlay punches through crushed cel shadow",
"shadow_tint": "#0b0b14",
"highlight_tint": "#ffd23f",
"grain": "clean digital cel + soft particle-glow bloom, no photographic grain",
"dynamic_range": "high, HDR compositing range on the light-FX layer",
"lut_match": "modern anime compositing grade: neutral cel base + saturated elemental light overlay"
}In a prompt this becomes: Color grade: saturation ~80/100, very high contrast…, shadow tint #0b0b14, highlight tint #ffd23f, grain: … — never "cinematic color grade".
Rule: 1 base (60–70%) + 1 accent (30–40%), the accent isolated on ONE element or surface. The base owns optics, WB and fps; the accent lends light or texture.
| Recipe | Base | Accent | Weights | Why it works |
|---|---|---|---|---|
| GHIBLIPUNK | GHIBLI | AKIRA | 70/30 | natureza gouache + um elemento neon isolado |
| GLASSNOIR | NOIR | LIQUIDGLASS | 70/30 | mono clássico + uma superfície de vidro refrativo |
| NEONOIR | NOIR | WONGKARWAI | 70/30 | P&B com neon saturado só nas fontes de luz |
| INKVERSE | SPIDERVERSE | MIGNOLA | 65/35 | 3D hachurado + blocos de preto sólido |
| SATURDAYCEL | XEROX | SYNTHWAVE | 70/30 | linha xerox trêmula + grid/chrome retrô |
| WESNOIR | WES | NOIR | 70/30 | simetria planimétrica com luz de slats |
| ANALOGPIXEL | ANALOGDREAM | PIXELART | 75/25 | foto quente com HUD/sprites pixelados isolados |
| CELVERSE | UFOTABLE | SPIDERVERSE | 70/30 | cel moderno + halftone/misregistration nos impactos |
| CHROMEKEYNOTE | KEYNOTE | LIQUIDMETAL | 70/30 | produto minimalista com uma morfose de metal líquido |
| VHSDREAM | DREAMCORE | FOUNDFOOTAGE | 70/30 | espaço liminar com ruído de fita e timecode |
| BIOGHIBLI | GHIBLI | BIOLUMINESCENT | 70/30 | floresta pintada que brilha à noite |
| WEBTOONCLAY | WEBTOON | CLAYMORPHISM | 70/30 | painel vertical com objetos puffy tácteis |
Known conflicts: LAIKA+PIXAR (matéria vs. render); XEROX+KEYNOTE (linha trêmula vs. precisão); PIXELART+PLANETEARTH (grade vs. teleobjetiva); Qualquer mistura de 3+ estilos.
🧾 IP: artist and studio names are cultural anchors for visual DNA (technical traits) — not a license to reproduce characters, logos, scenes or trade dress. For commercial use, rely on the
dna_tags/prompt_corefields and keep the IP LOCK.
| Type | Engines |
|---|---|
| Video | seedance_2_5 (up to 30s, ~50 refs) · seedance_2_0 · veo_3_1 (native audio) · kling_3_0 (multishot/4K) · grok_imagine_video · sora_2 runway_gen_4_5 · luma_ray3 · minimax_hailuo · wan_2_x · ltx_2_x |
| Image | flux_2 (brand HEX) · gpt_image_2 (text/UI) · nano_banana (4K, ~14 refs) · seedream_5 · midjourney · grok_imagine_image |
| Audio | elevenlabs (VO/SFX/dub) · suno (score) · native SCELA audio |
| Platform / pipeline / LLM | higgsfield (Cinema Studio + Soul ID) · comfyui (node DAG, PuLID/IP-Adapter) · claude |
One-line decision rule: dialogue and rich foley → Veo · cheap multishot/4K → Kling · long clip with many refs → Seedance 2.5 · fast social with sound → Grok · full local control → ComfyUI + LTX/Wan · recurring identity → Soul ID or PuLID (never text description alone) · anime/2D → still in the style first, then i2v, never ask for "3D".
confidence: low = specs not verified here; confirm on the platform. Sora 2 is sunsetting — do not start new pipelines on it.
Social creators · Motion designers & editors · Agencies · Brand owners & e-commerce · Growth marketers · Advanced AI-creative geeks · EdTech · Food · Fashion & beauty · UX/UI · Real estate · Filmmakers · Music · Games/anime/comics · Documentary & faceless · Cultural institutions · Independent animation, anime & webtoon (new) · Art direction & signature look (new).
See profiles_guide.md for pains, KPIs, default styles and quick-start commands.
The studio is Non-IP by default (Law 11, Skill 56, G8): no real brands, trade dress, recognizable voices or scores.
When a request cites a real brand or product, the engine neither refuses nor ignores it — it triggers G9 — Brand Mode, a consent gate with 5 questions:
- Which exact brand and product?
- Relationship: affiliate · sponsored · gifted · no relationship?
- Do you have the link (if affiliate) and accept labeling it in the description?
- Does the scene context fit the brand?
- Organic or paid ad (paid = written authorization)?
Only with all 5 answers does the mode turn on. Otherwise it continues Non-IP with a generic substitute (e.g. "classic black-and-white canvas sneakers, worn rubber sole"). Even with G9 on, these stay out: real people's face/voice without consent, recognizable score, third-party characters. Disclose whenever there is commission, gifted product or sponsorship. The disclosure label (#publi, #afiliado) does not count among the 3 hashtags. Not legal advice — rules vary by country and platform (FTC, CONAR, EU).
| Gate | What it locks |
|---|---|
| G1 | Style Bible: alias + HEX palette + optics + grading card before generating |
| G2 | Hero Frame approved in a still before animating |
| G3 | Feasibility Veto: no shot violates the engine's physics |
| G4 | Clean lint: no empty terms, unjustified mm, flat negatives, copy outside the copy list |
| G5 | Acting audit with <2 symptoms (15 AI symptoms: frozen blink, rubbery limbs, drifting eyelines…) |
| G6 | Audio/lip-sync: 1 line at a time, ≤25 words/10s, no loops |
| G7 | Darkroom: stochastic grain on every upscale, halation, tested LUT |
| G8 | Compliance: Non-IP, consent, synthetic labeling, 3 hashtags, 12%/15% safe zones |
| G9 | Brand Mode (opt-in, off by default) |
Feasibility Veto (examples): chaotic action + no blur · micro-acting at FOV ≥94° · 2+ dominant camera devices · on-screen text >6 words in a video engine · crowd with >5 individually acting faces · impossible single-take transformation (except declared LIQUIDMETAL) · 3+ styles in one generation.
Lint: errors NO_BLUR · DURATION · SPEECH_LEN · ENGINE_SUNSET · BRAND_UNGATED · BLEND_OVERLOAD — warnings EMPTY_TERM · MM_NOTATION · NEGATIVE_PHRASING · FOV_UNIT · GRADE_VAGUE · HASHTAGS · RATIO · BRAND_UNLISTED · BRAND_NO_DISCLOSURE.
import { resolveStyle, routeTask, recommendEngines, compilePrompt, lintPrompt, profileKit, blendStyles, gradeCard } from './skillsData';
resolveStyle('WES') // → VisualStyle with optics, HEX, prompt_core and color_grading resolved
routeTask('I want to animate this') // → route + skills + template + gate
recommendEngines({ kind: 'video', styleAlias: 'KUBRICK', durationS: 12 }) // → engines with verified_on and confidence
profileKit('perfil_18') // → skills (with deps), styles, engines and pipeline of the profile
blendStyles('GHIBLI', 'AKIRA', 0.7) // → merged optics/palette/prompt_core/color_grading + conflict flags
gradeCard('UFOTABLE') // → color_grading_card.json matched to a LUT preset
compilePrompt({ engineId: 'seedance_2_5', blend: { base: 'NOIR', accent: 'LIQUIDGLASS' }, subject, setting, action, camera })
lintPrompt(prompt) // → errors and warnings before you generateRequirement: tsconfig with "resolveJsonModule": true (Vite/Next already ship it). Verified with tsc --strict.
- New style: entry in
styles[]with a unique alias, an existingfamily,optics.fov_degrees, HEX,prompt_coreand a completecolor_grading— the.mdfiles and the TS update themselves - New blend recipe: row in
tables.ai_style_mixes.recipes - New engine: in
models[]withverified_on,confidence,grammar+family_routing - Brand watchlist:
tables.brand_mode.watchlist(lint trigger) - Stale specs:
staleEngines()lists what needs re-verification (90-day cycle)
- 3 hashtags per post (brand + niche + format) — the disclosure label is separate
- Packages: Essential (3/wk) · Performance (4–5/wk, A/B tests) · Premium (custom)
- CTR × Retention matrix: Winner → replicate the mechanic · Oversold by the cover → rewrite the body · Undervalued → new cover/title · Double failure → archive the mechanic
- Iteration hierarchy: thumbnail → title → hook (0–10s) → first open loop → body → closer
Daniel Rodrigues · Draft Creative Studio Ltd CineOS 3.2.0-cia · Cinematic Intelligence Architecture · Color Grading DNA · Style Blending · G9 Brand Mode
Idea, architecture and creative content: all rights reserved to the author. Contributions and bug reports are welcome via Issues.
From premise to post — with FOV in degrees and numbers in the grade, not empty adjectives. 🎬
— Daniel Rodrigues