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 preview images (OG 1200x630 + AI search optimization), stock photo curation (Pexels, P
Install
npx skills add https://github.com/heymegabyte/agent-skills/tree/main/12-media-orchestration
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install heymegabyte-agent-skills@llmmart
git clone https://github.com/heymegabyte/agent-skills.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole heymegabyte/agent-skills collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
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. Perplatform.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:
- Page topic + intent verbatim from
topic_intent - Brand palette tokens from
_brand.json.colors(inline hex) - Composition + aspect ratio matching
aspect - Subject specificity (NEVER "people" — always "octogenarian volunteer plating soup, soft window light, documentary style")
- Photographic technical specs (camera, lens, lighting, DoF — "shot on Hasselblad, 85mm prime, golden hour, shallow DoF")
- 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.
Files (agent-skills)
-
templates
-
PROMPTS.md 10 KB
Loading…
-
-
30-ideogram-methods.md 16 KB
Loading… -
build-breaking-rules.md 48.2 KB
Loading… -
compression-pipeline.md 3.8 KB
Loading… -
image-optimization.md 6.6 KB
Loading… -
image-profiling.md 3.2 KB
Loading… -
lightbox-classifier.md 10 KB
Loading… -
media-prompts.md 3.8 KB
Loading… -
notebooklm-pipeline.md 22.1 KB
Loading… -
og-card-pipeline.md 19.5 KB
Loading… -
og-image-generation.md 8.5 KB
Loading… -
SKILL.md 6.7 KB
--- name: "media-orchestration" description: "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 preview images (OG 1200x630 + AI search optimization), stock photo curation (Pexels, Pixabay), 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: - 30-ideogram-methods.md - build-breaking-rules.md - compression-pipeline.md - image-optimization.md - image-profiling.md - lightbox-classifier.md - media-prompts.md - notebooklm-pipeline.md - og-image-generation.md - social-brand-hex.md - technical-diagramming.md priority: 3 pack: "media" stage: stable triggers: - "image gen" - "dalle" - "media" - "og image" paths: - "org:website_build" --- # 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. -
social-brand-hex.md 3.8 KB
Loading… -
technical-diagramming.md 4.4 KB
Loading…
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.