--- name: "media-orchestration" description: "Section-by-section media planning and generation. Image generation (GPT Image 1.5 primary, built-in fallback), logo/icon generation (Ideogram v3 → favicon set), video generation (Sora), social preview images (OG 1200x630 + AI search optimization), stock photo curation (Unsplash, Pexels), critique/remix loops (max 3 rounds), asset compression pipeline, and media performance budgets." metadata: version: "2.1.0" updated: "2026-05-03" context: "fork" effort: "high" model: "sonnet" license: "Rutgers" compatibility: claude-code: ">=2.0.0" agentskills: ">=1.0.0" submodules: - media-prompts.md - compression-pipeline.md - og-image-generation.md - image-optimization.md - technical-diagramming.md - image-profiling.md - lightbox-classifier.md - social-brand-hex.md - notebooklm-pipeline.md - build-breaking-rules.md --- # 12 -- Media Orchestration Submodules: media-prompts.md (prompt templates, Ideogram v3 API), compression-pipeline.md (Python code, format tables, CF Image Transforms, CLS, broken image detection), og-image-generation.md (Satori edge-rendered OG images, KV/R2 cache, meta-tag helper), image-optimization.md (Sharp processing, responsive srcset, WebP/AVIF, blur placeholders, R2 pipeline), image-profiling.md (GPT-4o vision batch profiling — quality+placement+colors per image, pre-digest for builders), lightbox-classifier.md (per-image eligibility — kind!=logo + ≥1024×768 + score≥7, logo grids → hover-grayscale-to-color), social-brand-hex.md (canonical brand-color map per social platform, hover/focus/active states, per-platform CSS class generation), notebooklm-pipeline.md (per-site podcast via ElevenLabs Studio + infographic gallery via Vega-Lite/Recraft/GPT-Image-2 + explainer video via HeyGen + Cloudflare Stream embed + RSS feed + JSON-LD + cost ceiling $3.50/site, runs Phase 0 in parallel with Media Slot Manifest, paired with `notebooklm-orchestrator` agent). ## Strategy by Section Hero: GPT Image 1.5/Sora. Features: GPT Image 1.5/SVG. How It Works: GPT Image 1.5. Testimonials: stock. About: stock/real. Blog: GPT Image 1.5. Social: Satori OG 1200x630. Icons: Ideogram v3+processing. Pre-gen: communication goal? Brand style? Dimensions? Format? Budget? Stock or generated? ## Visual Inspection (MANDATORY) Read every image before deploy. Check: blur, artifacts, watermarks, wrong colors, AI hallucinations, gibberish text. Failed: regenerate improved prompt. Quality: 2x retina, no artifacts, brand palette, consistent style, no uncanny valley. ## Brian's Style Space/cosmic: #00E5FF + #7C3AED, deep black (#060610). Connections/dots: quantum, neural, constellation. "Ultra realistic" scenes. Transparent logos. Simpler always. Motifs: squirrels, turtles. ## Image Generation GPT Image 1.5 preferred (best quality). GPT Image 1 for speed. GPT Image 1-mini for bulk/drafts. Fallback: scripts/image_gen.py. Be specific, include colors, specify avoidances. Product screenshots: browser rendering via Playwright on live URL. ## DALL-E First Slot-Fill (***BRIAN'S STATED PREFERENCE — CANONICAL — UNIVERSAL***) DALL-E (gpt-image-1.5 / gpt-image-1 — colloquially "DALL-E") is the PRIMARY originator for every image slot the source-resolution chain didn't naturally fill from real-entity sources (Places/uploads/scrape). After real-entity sources exhaust, DALL-E is invoked BEFORE generic stock — per-slot prompt produces tighter topic match than any stock library. Stock APIs run as parallel speed-pass fallback (instant return if DALL-E hangs >15s) but DALL-E output preferred at curation. Quote: "DALL-E can literally create the ultra-realistic perfect photo for any given photo spot, so rely on that fact" — encoded as DEFAULT BEHAVIOR, not opt-in. See skill 15 media-acquisition Media-Slot-Manifest + Fail-CLOSED auto-regenerate (5 attempts, prompt-refinement loop, $0.40 worst-case ceiling per slot). ## Per-Slot Prompt Mandatory Fields (***UNIVERSAL — BUILD-BREAKING — validate-image-prompts.mjs + validate-dalle-slot-fill.mjs***) Every DALL-E call MUST encode 6 fields in the prompt, drawn from the slot record in `_media_slots.json`: (1) page topic + intent verbatim from `topic_intent`, (2) brand palette tokens from `_brand.json.colors` (inline hex), (3) composition + aspect ratio matching `aspect`, (4) subject specificity (NEVER "people" — always "octogenarian volunteer plating soup, soft window light, documentary style"), (5) photographic technical specs (camera, lens, lighting, depth of field, e.g. "shot on Hasselblad, 85mm prime, golden hour, shallow DoF"), (6) negative prompt block ("no text, no watermarks, no logos, no extra fingers, no AI artifacts, no stock-photo cliches"). Generic "create a hero image" prompts FAIL the validator. Same template applies to FLUX, Recraft, Stability — reuse slot-prompt across providers with fallback chain. ## Fail-CLOSED Auto-Regenerate (***ZERO MISSING IMAGES — UNIVERSAL — BUILD-BREAKING — validate-no-empty-slots.mjs***) Every slot in `_media_slots.json` MUST end the build with `filled_url != null AND filled_score >= relevance_floor` (default 8/10 via GPT-4o vision). Failure modes (Pexels returns nothing, DALL-E returns NSFW-flagged, scraped image broken, vision relevance below floor) trigger immediate auto-regeneration via DALL-E with REFINED prompt — NEVER silent skip, NEVER substitute brand-gradient unless 5 regen attempts exhausted. Refined prompt synthesized via gpt-4o reading `(original_prompt, vision_critique_of_what_was_wrong, relevance_floor)` and tightening with what to ADD + REMOVE. Hard ceiling 5 attempts × $0.08/img = $0.40/slot worst case. After exhaustion: log to `_unfillable_slots.json`, ship with brand-gradient floor (prevents 404), surface in dashboard for manual replacement, build marked `published_with_warnings`. Default behavior is regen-until-pass. **"DALL-E" terminology reconciliation:** when Brian or skill prose says "DALL-E", treat it as colloquial shorthand for the current OpenAI Images API (gpt-image-1.5 quality / gpt-image-1 balanced / gpt-image-1-mini bulk). The literal `dall-e-3` model id deprecated May 12, 2026 — never call that endpoint. Implementation always uses gpt-image-1.5 by default; the user-facing concept stays "DALL-E" because that's the brand language in user feedback. Sora video deprecates Sept 24, 2026. **Per-slot purpose-crafted prompt requirement (***UNIVERSAL — BUILD-BREAKING — no generic prompts***):** every generated image gets a per-slot prompt that names: (1) route+section the image lives in (`/about hero`, `/services card-3`, `/blog/post-x feature`), (2) page topic+intent (subject specificity — "octogenarian volunteer plating soup, soft window light, documentary realism" NOT "people helping"), (3) brand palette tokens (e.g. "matte navy #060610 + cyan accent #00E5FF"), (4) composition+aspect ratio (16:9 hero / 1:1 card / 4:5 portrait / 3:2 editorial), (5) negative prompt (no text overlays, no watermarks, no extra fingers, no logos, no AI artifacts), (6) stylistic anchor (matched to brand-element extraction — "documentary photography" / "minimal editorial" / "warm golden-hour"). Generic prompts ("create a hero image for nonprofit") rejected at the prompt-prep step. Same per-slot prompt template applies across providers — gpt-image-1.5, Ideogram, Stability, Recraft — fallback chain reuses the slot prompt verbatim. Validator (`validate-image-prompts.mjs`): every entry in `_assets.json.generated[]` must include `slot_prompt: { route, section, intent, palette, aspect, negative, style }` — missing fields=fail. **Hero media preference order (***UNIVERSAL — full-multimedia experience***):** every hero/page-banner section ranks media in this order: (1) original-source-hero asset preserved at ≥7/10 quality, (2) **Pexels Video API** HD loop (free, MP4 ≤2MB, 4-8s, autoplay+muted+loop+playsinline+poster), (3) Coverr/Mixkit free video, (4) gpt-image-1.5 generated still with brand-converged prompt, (5) Pexels/Unsplash photo, (6) brand-gradient with motif overlay (LAST RESORT). Video-over-still preference holds whenever ≥1 Pexels Video result scores ≥0.7 topical-relevance — sites should feel cinematic, not catalog. Ship `