media-orchestration

Section-by-section media planning and generation. Image generation (GPT Image 2 primary, built-in fallback), logo/icon generation (Ideogram v3 → favicon set), video generation (Veo 3.1), social previe

SKILL.md

GitHub

12 — Media Orchestration

Generate all site media section-by-section: images (GPT Image 2), logos (Ideogram v3), video (Veo 3.1), OG cards, and compression pipeline.

Model migration note (pass-77, 2026-06-09): DALL-E → GPT Image 2 (2026-04; 2.5 branded 2026-09 — API id unconfirmed, check current id at integration) + GPT-4o → GPT Image 2 vision. Per platform.openai.com/docs/deprecations.

Video rail (2026-10-04): Sora retired (app 2026-04-26, API 2026-09-24) — Veo 3.1 is the video rail. Fallbacks: Kling 3.0 / Runway Gen-4.5.

Submodules

  • media-prompts — prompt templates, Ideogram v3 API
  • compression-pipeline — Python code, format tables, CF Image Transforms, CLS, broken image detection
  • og-image-generation — Satori edge-rendered OG, KV / R2 cache, meta-tag helper
  • image-optimization — Sharp processing, responsive srcset, WebP/AVIF, blur placeholders, R2 pipeline
  • image-profiling — GPT Image 2 vision batch profiling
  • lightbox-classifier — per-image eligibility: kind!=logo + ≥1024×768 + score≥7
  • social-brand-hex — canonical brand-color map per social platform
  • notebooklm-pipeline — per-site podcast via ElevenLabs Studio + infographic via Vega-Lite/Recraft/GPT-Image-2 + HeyGen video + CF Stream + RSS + JSON-LD + cost ceiling $3.50/site

Strategy by Section

Hero → GPT Image 2 / Veo 3.1 · Features → GPT Image 2 / SVG · How It Works → GPT Image 2 · Testimonials → stock · About → stock/real · Blog → GPT Image 2 · Social → Satori OG 1200×630 · Icons → Ideogram v3

Pre-gen: communication goal · brand style · dimensions · format · budget · stock vs. generated?

Visual Inspection (MANDATORY)

Read every image before deploy: blur · artifacts · watermarks · wrong colors · hallucinations · gibberish text. Fail → regenerate with improved prompt. Quality bar: 2× 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 2 (gpt-image-2) preferred (best quality); GPT Image 1 for speed; GPT Image 1-mini for bulk/drafts
  • Fallback: scripts/image_gen.py; product screenshots: Playwright on live URL
  • Be specific: include colors, specify avoidances

GPT Image 2 First Slot-Fill (CANONICAL — UNIVERSAL)

PRIMARY originator for every slot real-entity sources (Places / uploads / scrape) didn't fill. GPT Image 2 invoked BEFORE generic stock; stock APIs run parallel speed-pass fallback (instant return if GPT Image 2 hangs >15s). See skill 15 media-acquisition + Fail-CLOSED auto-regenerate (5 attempts, $0.40 worst-case ceiling per slot).

Per-Slot Prompt Mandatory Fields (BUILD-BREAKING — validate-image-prompts.mjs + validate-dalle-slot-fill.mjs)

Every GPT Image 2 call MUST encode 6 fields from _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, DoF — "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 prompts FAIL validator. Same template applies to FLUX, Recraft, Stability.

Fail-CLOSED Auto-Regenerate (BUILD-BREAKING — validate-no-empty-slots.mjs)

Every slot MUST end build w/ filled_url != null AND filled_score >= relevance_floor (default 8/10 via GPT Image 2 vision). Failure modes (Pexels empty, NSFW-flagged, broken scrape, vision below floor) → immediate re-gen w/ REFINED prompt — NEVER silent skip, NEVER substitute brand-gradient unless 5 attempts exhausted. media_pipeline_orchestrator sub-agent owns this loop. See media-acquisition.md § Fail-CLOSED chain.

Logo / Icon / Video / OG

  • Logo — Ideogram v3 (best text rendering); Icons — Recraft V3; output: PNG transparent + SVG; bg removal → favicon set (16/32/180/192/512 + maskable); brand mark MUST be vector-clean
    • Ideogram 4.0 (open-weight, Apache 2.0) is current — generates native transparent backgrounds (reduces bg-strip need); v3 endpoint remains valid
  • Video — Veo 3.1 (primary rail: cinematic + narrative stitching, 7-8 × 8-sec clips → 60-sec arc); Kling 3.0 / Runway Gen-4.5 fallbacks; HeyGen (explainer/spokesperson); captions VTT + transcript; prefers-reduced-motion → static poster fallback
  • OG (1200×630) — Satori edge-rendered, per-route unique, BRANDED CARD never raw photo, ≤100KB, cached KV 7d / R2 forever

Stock Photography + Asset Compression + Performance

Stock: Pexels first (free, API); Pixabay second. Never Unsplash (generic), iStock/Getty (paid). Critique-and-remix loop max 3 rounds; AI vision <7/10 = reject + regenerate.

Compression: AVIF primary (94% browser support, 20-30% smaller than WebP); WebP fallback (Safari 14+); JPEG legacy. Sharp: 320/640/1280/1920w srcset; blur placeholder; dominant color → CSS bg fill. R2 upload per-extension content-type.

Budgets: total images/page ≤500KB; largest single ≤200KB; Hero LCP fetchpriority="high" + preload; loading="lazy" + decoding="async" on all others.

See submodules for: media-prompts, compression-pipeline, og-image-generation, image-optimization, image-profiling, lightbox-classifier, social-brand-hex, notebooklm-pipeline, build-breaking-rules.