cue-ai 0.9.2 → 0.9.3

package/resources/skills/skills/media/selfie-with-celebrities/SKILL.md

@@ -0,0 +1,64 @@
+---
+slug: muapi-selfie-with-celebrities
+name: muapi-selfie-with-celebrities
+version: "1.0.0"
+description: Generate a realistic behind-the-scenes selfie of the user with a celebrity or main actor from a specific movie, followed by an option to generate a cinematic long-take video connecting multiple selfies.
+acceptLicenseTerms: true
+---
+
+
+# Selfie with Celebrities
+
+**Generate a realistic behind-the-scenes selfie of the user with a celebrity or main actor from a specific movie, followed by an option to generate a cinematic long-take video connecting multiple selfies.**
+
+## Inputs
+
+| Name | Type | Required | Default | Description |
+|:---|:---|:---|:---|:---|
+| `movie_name` | text | yes | — | The name of the movie featuring the celebrity you want to take a selfie with (e.g., "Harry Potter", "The Matrix"). |
+| `user_image` | image_url | yes | — | A clear photo of the user to be used for the selfie. |
+
+
+## Steps
+
+### Phase A — Generate First Selfie Image
+
+If the `{{user_image}}` is not provided, ask the user to upload their image first.
+
+Once provided, submit the plan with ONE step:
+
+1. **Selfie Image Generation** — `muapi image edit` (model=`nano-banana-2-edit`):
+   - Reference Image: `{{user_image}}`
+   - Prompt: `POV: close up shot, A realistic photo of the person (don't change the person's original clothes) holding a black iPhone 16 to take a picture with the main actor from the movie "{{movie_name}}" on a movie set with movie scene background. Behind-the-scenes atmosphere, including film equipment, film cameras, filled with props, and busy crew members. Sharp focus on the characters. bright scene. Aspect ratio 9:16.`
+   - Aspect ratio: 9:16
+
+After generating the first image:
+- Present the generated selfie to the user.
+- Ask the user if they would like to create a second, third, etc. image with different movies or actors.
+- Also, suggest to the user that once they have multiple images, you can create a seamless cinematic video transitioning between the selfies.
+
+### Phase B — Generate Connecting Video (Only when requested)
+
+If the user has generated at least two selfies and asks to create a video connecting them, submit the plan with ONE step:
+
+1. **Cinematic Video Generation** — Use an image-to-video model like `kling-o1-image-to-video`, `pixverse-v5.5-i2v`, or `veo3.1-image-to-video`.
+   - First frame image: The first generated selfie.
+   - Last frame image: The second generated selfie (or use multiple steps if connecting more than two).
+   - Prompt: `A seamless cinematic long-take. The camera follows a person as she finishes a photo with an actor, then she naturally turns and walks toward the right. The camera tracks her movement with a smooth gimbal-like motion. Her gait is consistent and confident. Upon entering a new behind-the-scenes movie set, she slows down and stops gracefully, raising her phone with a bright smile to take a selfie with two actors. The scene transition is natural and smooth, with a seamless connection.`
+   - Aspect ratio: 9:16
+
+After generation:
+- Present the video to the user.
+
+## Trigger Keywords
+
+`selfie with celebrity`, `movie star selfie`, `celebrity photo`, `take a picture with actor`, `behind the scenes selfie`
+
+
+---
+
+## Notes for the Executing Agent
+
+- This recipe is LLM-orchestrated: read each phase, gather any missing inputs from the user, then call `muapi` CLI commands. Use `muapi auth configure` first if `MUAPI_API_KEY` is unset.
+- For model IDs without a CLI alias yet, fall back to the raw endpoint via `curl -X POST https://api.muapi.ai/api/v1/<endpoint> -H "x-api-key: $MUAPI_API_KEY" -H 'content-type: application/json' -d '{...}'` and poll with `muapi predict wait <request_id>`.
+- Substitute `{{input_name}}` placeholders with the user's actual inputs before issuing each call.

package/resources/skills/skills/media/social-media-video/SKILL.md

@@ -0,0 +1,277 @@
+---
+slug: muapi-social-media-video
+name: muapi-social-media-video
+version: "1.0.0"
+description: Brand-aware social media video creator. Reads brand-identity.md, ICP.md, and messaging.md to write a post/storyboard, craft an optimized Seedance 2.0 Director prompt, generate reference frames with the best available image model, and produce platform-ready video.
+acceptLicenseTerms: true
+---
+
+# Social Media Video Creator
+
+**End-to-end pipeline: Brand Files → Storyboard → Reference Images → Seedance 2.0 Video.**
+
+Reads your brand identity, ICP, and messaging documents to produce on-brand social video — fully optimized for Seedance 2.0's instructional prompt grammar and your target platform.
+
+---
+
+## Agent Execution Protocol
+
+### Step 1 — Read Brand Files
+
+Before writing anything, the agent MUST read all available brand files. Look for them in the working directory or any `brand/` subdirectory:
+
+| File | What to extract |
+|:---|:---|
+| `brand-identity.md` | Visual style, color palette, tone, logo/product aesthetics, brand personality |
+| `ICP.md` | Target audience — who they are, their pain points, what motivates them |
+| `messaging.md` | Core value props, hooks, CTAs, campaign themes, taglines |
+
+If a file is missing, proceed with what's available and note the gap.
+
+---
+
+### Step 2 — Write the Social Post + Storyboard
+
+Use brand context to produce:
+
+**Social Post Copy** (for caption/copy):
+- Hook line (first 1–2 sentences — must stop the scroll)
+- Body (3–5 sentences: problem → solution → proof → CTA)
+- Hashtags (5–8 relevant tags)
+- CTA (one clear action)
+
+**Storyboard** (match duration to platform spec):
+```
+0–3s:  [Opening scene — hook/visual surprise — camera move]
+3–7s:  [Core message — product/subject in action — camera move]
+7–10s: [Resolution — brand moment — CTA text on screen]
+10–15s: [Logo/tagline hold — background music fade]
+```
+
+Tone must match brand personality. If brand is playful → upbeat transitions, bright grade.
+If brand is premium → slow reveals, dark luxury aesthetic, moody lighting.
+
+---
+
+### Step 3 — Craft the Seedance 2.0 Director Prompt
+
+Transform the storyboard into a **technical Director Brief** for Seedance 2.0.
+
+**Rules:**
+1. Never use vague descriptors ("beautiful", "amazing", "8k"). Use technical cinematography language.
+2. Always specify camera movement, lens type, and lighting physically.
+3. For 10s+ videos, use timecode segments: `0–3s: [...] 3–7s: [...] 7–10s: [...]`
+4. Integrate `@image1`, `@image2` reference tags if images are provided.
+5. Always include sound direction (even brief) — Seedance generates audio.
+6. Lead with composition, end with texture and micro-motion.
+
+**Director Brief Template:**
+```
+[SCENE] {environment, lighting, time of day}
+[SUBJECT] {product/character/subject with specific detail}
+[ACTION] {what happens — fluid, continuous, physically plausible}
+[CAMERA] {movement + lens + framing}
+[STYLE] {color grade, mood, film reference if helpful}
+[SOUND] {music tone, sound effects, ambient}
+0–Xs: {if multi-beat}
+```
+
+---
+
+### Step 4 — Generate Reference Images (If Needed)
+
+**When to generate reference images:**
+
+| Scenario | Mode | Images Needed |
+|:---|:---|:---|
+| Product showcase | `i2v` | 1 product shot as first frame |
+| Scene transition | `first-last` | 2 images — opening and closing frame |
+| Brand character | `i2v` | 1 character reference |
+| Pure concept | `t2v` | None — text only |
+| Mood/style anchor | `i2v` | 1 style reference image |
+
+**Image generation — best models by use case:**
+
+| Use Case | Recommended Model | Why |
+|:---|:---|:---|
+| Photorealistic product/scene | `google-imagen4-ultra` | Highest realism, great lighting |
+| Concept art / stylized | `flux-kontext-pro-t2i` | Creative fidelity, style adherence |
+| Fastest turnaround | `google-imagen4-fast` | Speed with good quality |
+| Highly detailed/editorial | `hidream-i1-full` | Fine detail, editorial quality |
+| Character with identity | `ideogram-v3-t2i` | Strong text + character rendering |
+
+**Reference image prompt format:**
+Write a clean, technical image prompt (not a Seedance prompt). Include:
+- Subject description + key product/brand visual elements
+- Lighting (studio, golden hour, etc.)
+- Shot framing (medium shot, product close-up, etc.)
+- Mood/color palette matching brand identity
+- NO motion language (this is for a still frame)
+
+Execute image generation:
+```bash
+bash core/media/generate-image.sh \
+  --model google-imagen4-ultra \
+  --prompt "your image prompt" \
+  --aspect-ratio 9:16 \
+  --view
+```
+
+---
+
+### Step 5 — Generate the Video
+
+Choose mode, tier, and camera based on content type and available assets.
+
+**Mode selection:**
+
+| Situation | Mode | Command |
+|:---|:---|:---|
+| No reference images | `t2v` | default |
+| 1 image (first frame) | `i2v` | `--mode i2v --file ref.jpg` |
+| Start + end frames | `first-last` | `--mode first-last --tier global --file start.jpg --file end.jpg` |
+| Multi-ref blend | `i2v` | up to 9 images |
+
+**Invoke the script:**
+```bash
+bash library/social/social-media-video/scripts/run-social-video.sh \
+  --prompt "your director brief here" \
+  --platform instagram \
+  --camera drone \
+  [--mode t2v|i2v|first-last] \
+  [--file ref_image.jpg] \
+  [--gen-ref "reference image prompt"] \
+  [--tier global] \
+  [--quality high] \
+  [--view]
+```
+
+---
+
+## Platform Specs
+
+| Platform | Format | Aspect | Duration | Notes |
+|:---|:---|:---|:---|:---|
+| Instagram Reels | Vertical | 9:16 | 10–15s | Hook in first 1s |
+| Instagram Feed | Square | 1:1 | 10s | Static-feel works well |
+| TikTok | Vertical | 9:16 | 10–15s | High energy, fast cuts |
+| YouTube Shorts | Vertical | 9:16 | 15s | Max quality |
+| LinkedIn | Landscape | 16:9 | 10–15s | Professional tone |
+| Twitter/X | Landscape | 16:9 | 10s | Punchy, direct |
+| YouTube (long) | Landscape | 16:9 | 15s | Cinematic, slow builds |
+| Pinterest | Portrait | 4:3 | 10s | Lifestyle-forward |
+
+> **Tier note:** Use `--tier global` or `--tier vip` for `1:1` and `21:9` formats. Chinese tier supports only 16:9, 9:16, 4:3, 3:4.
+
+---
+
+## Camera Language Reference
+
+### Standard Camera Intents (--camera / --intent flags)
+
+| Intent | Movement | Best For |
+|:---|:---|:---|
+| `reveal` | Slow crane up, wide establishing | Product launches, brand reveals |
+| `epic` | Dolly in + orbit, low hero angle | Brand manifesto, emotional story |
+| `product` | Static macro orbit, precision reveal | E-commerce, product demo |
+| `narrative` | Tracking shot, Steadicam | Testimonials, story-driven |
+| `tense` | Handheld jitter, dutch angle | High-energy, urgency |
+| `comedy` | Reactive handheld, punchy zooms | Lighthearted brand content |
+
+### Specialty Camera Intents (New)
+
+| Intent | Description | Best For |
+|:---|:---|:---|
+| `fpv` | First-person subjective POV — immersive GoPro-style, continuous forward motion, peripheral detail close-ups | Action brands, travel, sports, tech demos |
+| `drone` | Aerial cinematic flythrough — smooth gimbal-stabilized, sweeping laterals, descend from high altitude into scene | Real estate, luxury, outdoor brands, epic reveals |
+| `flythrough` | Ground-level architectural flythrough — continuous dolly through space, seamless portal transitions | Architecture, interior design, venue showcases |
+
+**FPV Prompt Enrichment:**
+```
+Immersive first-person POV shot. Camera glides forward through [scene] at eye level.
+Slight natural motion stabilization with GoPro-style wide angle.
+Peripheral detail rushing past — [details]. Smooth continuous forward motion.
+No cuts throughout. [Subject] visible in foreground periphery.
+```
+
+**Drone Flythrough Prompt Enrichment:**
+```
+Cinematic aerial drone shot. Camera descends from 200m altitude toward [subject/scene],
+sweeping lateral arc as it descends. Gimbal-stabilized smooth motion.
+Golden hour atmosphere, long shadows across [terrain/scene].
+Final frame settles into medium establishing shot.
+Aerial cinematography, DJI Inspire aesthetic.
+```
+
+---
+
+## Prompt Quality Checklist
+
+Before finalizing the Seedance prompt, verify:
+
+- [ ] Scene environment is physically specific (not "nice background")
+- [ ] Camera movement named explicitly (dolly in, orbit, drone flythrough, FPV, etc.)
+- [ ] Lighting described technically (volumetric god rays, rim lighting, soft diffused, etc.)
+- [ ] Subject/product described with visual specifics from brand-identity.md
+- [ ] Sound direction included (even one line)
+- [ ] Timecodes used for 10s+ videos
+- [ ] `@image1` etc. referenced if images are provided
+- [ ] Brand CTA or tagline included in final seconds
+- [ ] No vague adjectives ("amazing", "beautiful", "stunning") — replaced with technical terms
+
+---
+
+## Example: Full Workflow
+
+**User**: "Make an Instagram Reel for our cold brew coffee brand, drone shot, premium feel"
+
+**Step 1 — Brand read**: Read `brand-identity.md` (minimalist packaging, dark roast, black + gold palette), `ICP.md` (urban professionals 25–40, values quality), `messaging.md` ("Precision Brewed. Zero Compromise.")
+
+**Step 2 — Storyboard**:
+```
+0–2s:  Drone descends over rooftop terrace at sunrise, fog below.
+2–5s:  Drone swoops down toward coffee cup on white marble, steam rising.
+5–9s:  Close-up orbit of bottle, specular gold highlights, ice cubes.
+9–11s: Product settles, black screen fades in: "Zero Compromise." + logo
+Sound: Minimal lounge beat, coffee pour sound effect at 5s.
+```
+
+**Step 3 — Seedance Director Prompt**:
+```
+0–2s: Cinematic aerial drone shot. Camera descends at 30° angle toward a rooftop terrace at sunrise.
+Golden hour atmosphere, San Francisco bay fog below horizon. Gimbal-stabilized smooth descent.
+2–5s: Drone rapidly closes toward a glass of cold brew coffee on white marble.
+Steam curling upward. Macro lens approach. Caustic light patterns on wet glass surface.
+5–9s: Slow precision orbit around cold brew bottle. Black matte label with gold embossed text catching
+specular highlights. Ice cubes with subsurface light scattering. Commercial macro aesthetic.
+9–11s: Static product hero shot. Letterbox crop, deep focus, black fade in from sides.
+Sound: Minimal ambient beat, single piano note, coffee liquid sound effect at 5s mark.
+Maintain cinematic color grade — deep blacks, warm gold midtones throughout.
+```
+
+**Step 4 — Generate reference**: `google-imagen4-ultra` → cold brew bottle product shot (9:16)
+
+**Step 5 — Generate video**:
+```bash
+bash library/social/social-media-video/scripts/run-social-video.sh \
+  --prompt "0–2s: Cinematic aerial drone shot..." \
+  --platform instagram \
+  --camera drone \
+  --mode i2v \
+  --file media_outputs/coldbrew_ref.jpg \
+  --duration 11 \
+  --tier global \
+  --view
+```
+
+---
+
+## Common Mistakes to Avoid
+
+1. **Reading brand files and ignoring them** — the storyboard must visually match the brand palette and tone.
+2. **Generic prompts** — "a nice video of a product" produces generic output. Every token must direct.
+3. **Wrong tier for aspect ratio** — 1:1 and 21:9 require `--tier global` or `--tier vip`.
+4. **Forgetting sound** — Seedance generates audio. Direct it, or you get random results.
+5. **FPV with static subject** — FPV requires continuous motion in the scene. Pair with movement-rich environments.
+6. **Drone without establishing shot** — drone works best when it resolves INTO something (a product, a scene, a subject).
+7. **Too many scene changes in 5s** — match complexity to duration. 5s = 1 beat. 15s = 3–4 beats.

package/resources/skills/skills/media/social-media-video/scripts/run-social-video.sh

@@ -0,0 +1,316 @@
+#!/bin/bash
+# Social Media Video Generator — Orchestrator
+#
+# Generates reference images (if requested) then produces a Seedance 2.0 video,
+# all wired to platform-specific defaults (aspect ratio, duration, tier).
+#
+# Usage:
+#   bash run-social-video.sh --prompt "director brief" --platform instagram [options]
+#
+# Requires: bash 3.2+, curl, jq
+
+set -euo pipefail
+
+# ============================================================
+# Locate skills root (works regardless of CWD when invoked)
+# ============================================================
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SKILLS_ROOT="$(cd "$SCRIPT_DIR/../.." && pwd)"
+GENERATE_IMAGE="$SKILLS_ROOT/core-media/generate-image.sh"
+GENERATE_SEEDANCE="$SKILLS_ROOT/seedance-2/scripts/generate-seedance.sh"
+
+# Load .env from skills root
+if [ -f "$SKILLS_ROOT/.env" ]; then
+    set +u; source "$SKILLS_ROOT/.env" 2>/dev/null || true; set -u
+fi
+
+# ============================================================
+# Defaults
+# ============================================================
+PROMPT=""
+PLATFORM="instagram"
+CAMERA=""
+MODE="t2v"
+TIER="chinese"
+FAST=false
+ASPECT=""
+DURATION=""
+QUALITY="basic"
+VIEW=false
+ASYNC=false
+IMAGE_FILES=()
+
+# Reference image generation
+GEN_REF_FIRST=""      # --gen-ref "prompt"       → generate first frame ref
+GEN_REF_LAST=""       # --gen-ref-last "prompt"  → generate last frame ref
+REF_MODEL="google-imagen4-ultra"
+
+# ============================================================
+# Argument parsing
+# ============================================================
+while [[ $# -gt 0 ]]; do
+    case $1 in
+        --prompt|-p)        PROMPT="$2";         shift 2 ;;
+        --platform)         PLATFORM="$2";       shift 2 ;;
+        --camera)           CAMERA="$2";         shift 2 ;;
+        --mode)             MODE="$2";           shift 2 ;;
+        --tier)             TIER="$2";           shift 2 ;;
+        --fast)             FAST=true;           shift   ;;
+        --aspect)           ASPECT="$2";         shift 2 ;;
+        --duration)         DURATION="$2";       shift 2 ;;
+        --quality|-q)       QUALITY="$2";        shift 2 ;;
+        --view)             VIEW=true;           shift   ;;
+        --async)            ASYNC=true;          shift   ;;
+        --file|-f)          IMAGE_FILES+=("$2"); shift 2 ;;
+        --gen-ref)          GEN_REF_FIRST="$2";  shift 2 ;;
+        --gen-ref-last)     GEN_REF_LAST="$2";   shift 2 ;;
+        --ref-model)        REF_MODEL="$2";      shift 2 ;;
+        --help|-h)
+            cat <<'HELP'
+Social Media Video Generator
+Usage: bash run-social-video.sh --prompt "director brief" [options]
+
+REQUIRED
+  --prompt, -p TEXT     Seedance 2.0 Director Brief (the full cinematic prompt)
+
+PLATFORM (sets aspect/duration defaults)
+  --platform NAME       instagram | tiktok | linkedin | youtube | youtube-short |
+                        twitter | pinterest | instagram-feed  (default: instagram)
+
+CAMERA STYLE (appends camera directive to prompt)
+  --camera TYPE         fpv | drone | flythrough | reveal | epic | product |
+                        narrative | tracking | orbit  (default: from --platform)
+
+VIDEO OPTIONS
+  --mode MODE           t2v | i2v | first-last | omni  (default: t2v)
+  --tier TIER           chinese | global | vip  (default: chinese)
+                        Note: first-last requires global or vip
+  --fast                Use fast-queue variant (global/vip only)
+  --aspect RATIO        Override platform default aspect ratio
+  --duration N          Override platform default duration (seconds)
+  --quality Q           basic | high  (Chinese tier only; default: basic)
+  --view                Download and open output (macOS)
+  --async               Return request_id without waiting
+
+REFERENCE IMAGES
+  --file, -f PATH       Provide an existing image file (repeatable, up to 9)
+  --gen-ref TEXT        Generate a first-frame reference image from this prompt
+  --gen-ref-last TEXT   Generate a last-frame reference image (triggers first-last mode)
+  --ref-model MODEL     Image model for reference generation
+                        (default: google-imagen4-ultra)
+                        Options: flux-kontext-pro-t2i | hidream-i1-full |
+                                 google-imagen4-fast | ideogram-v3-t2i
+
+EXAMPLES
+  # Instagram Reel — drone shot, auto-generate ref image, t2v
+  bash run-social-video.sh \
+    --prompt "Drone descends over rooftop at sunrise, product reveal..." \
+    --platform instagram --camera drone --duration 10
+
+  # TikTok — FPV, with generated first frame
+  bash run-social-video.sh \
+    --prompt "First-person rush through..." \
+    --platform tiktok --camera fpv \
+    --gen-ref "Hero product on table, soft morning light, close-up" \
+    --tier global --view
+
+  # LinkedIn landscape — product orbit, existing reference image
+  bash run-social-video.sh \
+    --prompt "Precision orbit around product, specular highlights..." \
+    --platform linkedin --camera product --mode i2v --file product.jpg \
+    --quality high --view
+
+  # YouTube Short — first-last frame interpolation
+  bash run-social-video.sh \
+    --prompt "Dramatic reveal from foggy landscape to clear skyline..." \
+    --platform youtube-short \
+    --gen-ref "Morning fog over mountain valley, sunrise" \
+    --gen-ref-last "Clear blue sky, mountain peak, golden hour" \
+    --tier global --view
+HELP
+            exit 0 ;;
+        *)
+            echo "Warning: Unknown argument '$1' — skipping." >&2
+            shift ;;
+    esac
+done
+
+# ============================================================
+# Validate
+# ============================================================
+if [ -z "$PROMPT" ]; then
+    echo "Error: --prompt is required. Provide the Seedance Director Brief." >&2
+    exit 1
+fi
+if [ -z "${MUAPI_KEY:-}" ]; then
+    echo "Error: MUAPI_KEY is not set. Export it or add to $SKILLS_ROOT/.env" >&2
+    exit 1
+fi
+
+# ============================================================
+# Platform → aspect ratio + duration defaults
+# ============================================================
+case "$PLATFORM" in
+    instagram|tiktok|youtube-short)
+        DEFAULT_ASPECT="9:16"; DEFAULT_DURATION=10 ;;
+    instagram-feed)
+        DEFAULT_ASPECT="1:1";  DEFAULT_DURATION=10 ;;
+    linkedin|youtube|twitter)
+        DEFAULT_ASPECT="16:9"; DEFAULT_DURATION=10 ;;
+    pinterest)
+        DEFAULT_ASPECT="4:3";  DEFAULT_DURATION=10 ;;
+    *)
+        DEFAULT_ASPECT="16:9"; DEFAULT_DURATION=10 ;;
+esac
+
+ASPECT="${ASPECT:-$DEFAULT_ASPECT}"
+DURATION="${DURATION:-$DEFAULT_DURATION}"
+
+# Formats requiring global/vip tier
+if [[ "$ASPECT" == "1:1" || "$ASPECT" == "21:9" ]] && [[ "$TIER" == "chinese" ]]; then
+    echo "Note: Aspect ratio $ASPECT requires global or vip tier. Switching to global." >&2
+    TIER="global"
+fi
+
+# ============================================================
+# Camera → intent mapping (for generate-seedance.sh --intent)
+# ============================================================
+case "$CAMERA" in
+    fpv)                INTENT="fpv" ;;
+    drone|flythrough)   INTENT="drone" ;;
+    reveal)             INTENT="reveal" ;;
+    epic)               INTENT="epic" ;;
+    product)            INTENT="product" ;;
+    narrative|tracking) INTENT="narrative" ;;
+    orbit)              INTENT="epic" ;;
+    *)                  INTENT="cinematic" ;;
+esac
+
+# ============================================================
+# Generate reference images if requested
+# ============================================================
+OUTPUT_DIR="${PWD}/media_outputs"
+mkdir -p "$OUTPUT_DIR"
+
+generate_ref_image() {
+    local REF_PROMPT="$1"
+    local LABEL="$2"
+    local OUTPUT_PATH="$OUTPUT_DIR/social_ref_${LABEL}_$(date +%s).jpg"
+
+    echo "" >&2
+    echo "Generating $LABEL reference image with $REF_MODEL..." >&2
+    echo "Prompt: $REF_PROMPT" >&2
+
+    local RESULT
+    RESULT=$(cd "$SKILLS_ROOT" && bash "$GENERATE_IMAGE" \
+        --model "$REF_MODEL" \
+        --prompt "$REF_PROMPT" \
+        --aspect-ratio "$ASPECT" \
+        --json)
+
+    local IMG_URL
+    IMG_URL=$(echo "$RESULT" | jq -r '.outputs[0] // empty')
+
+    if [ -z "$IMG_URL" ]; then
+        echo "Error: Failed to generate reference image. Raw response:" >&2
+        echo "$RESULT" >&2
+        exit 1
+    fi
+
+    echo "Downloading reference image to $OUTPUT_PATH..." >&2
+    curl -s -o "$OUTPUT_PATH" "$IMG_URL"
+
+    if [ "$VIEW" = true ] && [[ "$OSTYPE" == "darwin"* ]]; then
+        open "$OUTPUT_PATH"
+    fi
+
+    echo "$OUTPUT_PATH"
+}
+
+# Generate first-frame reference
+if [ -n "$GEN_REF_FIRST" ]; then
+    FIRST_REF_PATH=$(generate_ref_image "$GEN_REF_FIRST" "first")
+    IMAGE_FILES=("$FIRST_REF_PATH" "${IMAGE_FILES[@]}")
+    echo "First-frame reference: $FIRST_REF_PATH" >&2
+fi
+
+# Generate last-frame reference
+LAST_REF_PATH=""
+if [ -n "$GEN_REF_LAST" ]; then
+    LAST_REF_PATH=$(generate_ref_image "$GEN_REF_LAST" "last")
+    echo "Last-frame reference: $LAST_REF_PATH" >&2
+fi
+
+# ============================================================
+# Determine mode from available images
+# ============================================================
+# If we have a last-frame reference, use first-last mode
+if [ -n "$LAST_REF_PATH" ]; then
+    MODE="first-last"
+    IMAGE_FILES+=("$LAST_REF_PATH")
+    if [ "$TIER" = "chinese" ]; then
+        echo "Note: first-last mode requires global/vip tier. Switching to global." >&2
+        TIER="global"
+    fi
+fi
+
+# If we have images and mode is still t2v, switch to i2v
+if [ ${#IMAGE_FILES[@]} -gt 0 ] && [ "$MODE" = "t2v" ]; then
+    MODE="i2v"
+fi
+
+# ============================================================
+# Build generate-seedance.sh command
+# ============================================================
+SEEDANCE_ARGS=(
+    --mode "$MODE"
+    --tier "$TIER"
+    --subject "$PROMPT"
+    --intent "$INTENT"
+    --aspect "$ASPECT"
+    --duration "$DURATION"
+)
+
+# Quality only applies to Chinese tier
+if [ "$TIER" = "chinese" ]; then
+    SEEDANCE_ARGS+=(--quality "$QUALITY")
+fi
+
+# Fast flag (global/vip only)
+if [ "$FAST" = true ]; then
+    SEEDANCE_ARGS+=(--fast)
+fi
+
+# Attach reference images
+for IMG in "${IMAGE_FILES[@]}"; do
+    SEEDANCE_ARGS+=(--file "$IMG")
+done
+
+# Pass-through flags
+[ "$VIEW" = true ]  && SEEDANCE_ARGS+=(--view)
+[ "$ASYNC" = true ] && SEEDANCE_ARGS+=(--async)
+
+# ============================================================
+# Summary
+# ============================================================
+echo "" >&2
+echo "======================================" >&2
+echo " Social Media Video Generator" >&2
+echo "======================================" >&2
+echo " Platform  : $PLATFORM" >&2
+echo " Aspect    : $ASPECT" >&2
+echo " Duration  : ${DURATION}s" >&2
+echo " Mode      : $MODE" >&2
+echo " Tier      : $TIER" >&2
+echo " Camera    : ${CAMERA:-default}" >&2
+echo " Intent    : $INTENT" >&2
+echo " Ref images: ${#IMAGE_FILES[@]}" >&2
+echo " Ref model : $REF_MODEL" >&2
+echo "======================================" >&2
+echo "" >&2
+
+# ============================================================
+# Execute Seedance generation
+# ============================================================
+cd "$SKILLS_ROOT"
+exec bash "$GENERATE_SEEDANCE" "${SEEDANCE_ARGS[@]}"