@azerogluemin/ai-bootstrap 0.2.1

package/templates/skills/video-prompt-engineer/knowledge/models/kling-2.md

@@ -0,0 +1,280 @@
+# Kling 2.x — Video Model Knowledge
+
+## Provider
+Kuaishou (China)
+
+## Type
+Diffusion
+
+## Strengths
+
+- **Realistic human motion** (best in class for walking, gesturing)
+- **Lip sync** is strongest among AI video models
+- More complex motion supported than Veo
+- Cinematic look with proper prompting
+- 10-second clips supported
+- Good with Asian features (and others)
+- Image-to-video conversion
+
+## Weaknesses
+
+- No audio dependency (must add audio in post)
+- Long takes lose consistency
+- Western faces sometimes "Asian-leaning"
+- Verbose prompts confuse the model
+- Documentation primarily Chinese
+
+## Best Practices
+
+1. **Structured key-value format** — not natural language
+2. Maximum 80-100 words (much shorter than Veo)
+3. Cinematic style MUST be explicitly stated
+4. Aspect ratio and duration as parameters, not prompt text
+5. Image input supported and recommended
+6. Lip sync mode for dialogue scenes
+
+## Prompt Anatomy
+
+```
+Subject: [description]
+Action: [motion description]
+Camera: [shot type, angle, movement]
+Setting: [environment]
+Lighting: [setup]
+Style: [cinematic notes]
+Duration: Ns
+```
+
+## Sample Prompt
+
+```
+Subject: 35-year-old man, charcoal wool coat, standing by window
+Action: slow head turn from right to left, looks toward door, no body movement
+Camera: medium shot, eye-level, completely static, no camera movement
+Setting: 1960s diner interior at night, rain visible through window
+Lighting: warm tungsten key from ceiling + cool blue window backlight
+Style: cinematic, melancholic mood, shallow depth of field, photorealistic
+Duration: 6s
+```
+
+## Lip Sync Mode
+
+For dialogue clips, Kling has dedicated lip sync. Provide:
+- Audio file (or text-to-speech reference)
+- Subject framed for face visibility
+- Static or minimal camera
+
+```
+Subject: woman, 30 years old, dark hair, professional attire
+Action: speaking the dialogue, natural facial expression matching emotion
+Camera: medium close-up, static, eye-level
+Lip sync to: "I never expected to see you here." [tone: surprised, soft]
+Duration: 5s
+```
+
+## Avatar Mode (Kling AI Avatar — image + audio → talking video)
+
+Researched 2026-06-15 from official sources: kling.ai, higgsfield, scenario, pollo.
+
+### Inputs (required)
+
+- **Image**: static character image (close-up or chest-up, front-facing, single subject)
+- **Audio**: clean voice file (the actual speech to be lip-synced — NOT a text dialogue tag like in 3.0 dialogue mode)
+- **Prompt**: short performance guidance (NOT full Veo-style scene description)
+
+### Image requirements (strict)
+
+- Front-facing or near-frontal (slight 3/4 OK)
+- Face well-lit, eyes open
+- **No heavy occlusions on the face** — hands at face level, microphones, sunglasses block lip-sync
+- Hands at chest level (selfie pose) are acceptable as long as the mouth is fully visible
+- Supports photoreal portraits, anime, illustration, stylized avatars
+
+### Audio requirements
+
+- Minimal background noise
+- Clear pronunciation
+- Conversational pace — extremely fast or slow speech sync poorly
+- Supports speech AND singing
+- Multi-language (works with Azerbaijani, Turkish, Russian, etc.)
+
+### Prompt structure (concise — NOT long)
+
+The prompt describes 3 things in order:
+
+1. **Persona** — who they are (age, brief context)
+2. **Expression** — concrete emotional descriptors (NOT "happy" → use "smiling warmly", NOT "angry" → use "brows furrowed, lips pressed tight")
+3. **Motion style** — gestures, head movement, body language, camera behavior, pace, language
+
+### Avatar prompt length (researched 2026-06-14 from official + production sources)
+
+**HARD finding**: Kling does NOT publish an explicit character/word limit for the Avatar prompt field. Avatar is officially documented as "optional text prompt to influence background style, mood, or framing." The IMAGE carries the scene, the AUDIO carries the speech — the prompt is performance guidance only.
+
+**Practical word-count guidance** (across all credible 2026 sources):
+
+| Range | Quality | Risk |
+|---|---|---|
+| **<15 words** | Vague, model defaults to generic motion | High under-instruction |
+| **15-40 words** | Optimal for image-to-video / Avatar (matches image-to-video sweet spot per VEED) | Low — production-safe |
+| **40-80 words** | Acceptable, but every additional sentence is potential contradiction | Low-medium |
+| **80-100 words** | Approaching ceiling — possible only if instructions don't conflict | Medium |
+| **>100 words** | High failure risk — contradictory instructions, ignored fields, generation hangs | HIGH |
+| **>150 words** | Strongly discouraged for Avatar/image-to-video mode | Very high |
+
+**Production rule of thumb**:
+- **Default target**: 40-60 words (persona + expression + motion + gestures + eyes + camera + language)
+- **Hard cap**: 100 words
+- **If close to 100**: cut redundancy — language tag, "static camera", and "pace" can be 1-2 words each
+
+**Why short matters for Avatar specifically**:
+- The IMAGE is the scene anchor — repeating scene details in the prompt is wasted tokens and risks contradicting the image
+- The AUDIO is the speech anchor — describing what they say is wasted; only describe HOW they perform it
+- Avatar Mode is performance direction, NOT scene-building (which is Veo/Sora territory)
+
+### Sources (Avatar prompt length, verified 2026-06-14)
+
+- [VEED — Kling AI Prompts Complete Guide 2026](https://www.veed.io/learn/kling-ai-prompting-guide) — image-to-video sweet spot 15-40 words (Avatar matches this category)
+- [Kling AI Avatar 2.0 User Guide — kling.ai](https://kling.ai/quickstart/kling-ai-avatar-2-user-guide) — official: "optional text prompt"
+- [VideoAI — Kling AI Prompt Length Guide](https://videoai.me/blog/kling-ai-prompt-length) — general production data, 100+ words = failure risk
+- [ReelMind — Kling AI Prompt Character Limit](https://reelmind.ai/blog/kling-ai-prompt-character-limit-ai-for-artistry-limits) — confirms no published hard limit
+
+### Avatar prompt anti-patterns (drop these to stay under 100 words)
+
+| Anti-pattern | Why drop |
+|---|---|
+| Full setting/scene description | The IMAGE already carries this — repeating wastes tokens |
+| Describing what they say | The AUDIO carries speech — prompt only describes performance |
+| Multiple example phrases ("on word X they do Y, on word Y they do Z") | The lip-sync engine handles word-level timing; high-level performance guidance is enough |
+| Repeating identity details (face, clothing) | The IMAGE locks identity — only mention persona context briefly (age, role) |
+| Long camera explanation | Avatar is static-camera by design — one line max ("static, 9:16, chest-up") |
+| Verbose language tag ("the audio is in Russian and Azerbaijani mixed market vendor pidgin which the lip-sync should handle smoothly") → "Russian-Azerbaijani mix" |
+
+### Avatar prompt template
+
+```
+A [persona — age, identifier, brief context], [setting].
+
+Expression: [concrete emotional descriptors — eyebrow movement, mouth shape on emphasis, micro-expressions on specific words].
+
+Motion style: [head tilts, body lean, breathing, energy level — animated/restrained/calm].
+
+Gestures: [hand position, micro-movements, whether hands stay raised or drop, gesture frequency].
+
+Eyes: [locked on camera / off-camera / shifting].
+
+Camera: [static / minimal pan / movement], [aspect ratio], [framing].
+
+Pace: [conversational / slow / fast — matching the audio rhythm].
+
+Language: [language of the audio for lip-sync].
+```
+
+### Avatar prompt example (real-world, AZ context)
+
+```
+A 45-50 year old Azerbaijani woman with a distinctive larger nose, in her home kitchen, on a video call.
+
+Expression: speaking confidently and assertively, slightly defiant, eyebrows lifted on emphasis, mouth fully articulating each Azerbaijani word, occasional knowing micro-smile on pointed lines.
+
+Motion style: animated but controlled — small natural head tilts and slight head shakes on rejecting phrases, subtle forward lean on emphasized declarations.
+
+Gestures: right hand stays raised at chest level (selfie-holding position) throughout, with small finger-pointing or open-palm emphasis gestures, never dropping.
+
+Eyes: locked directly on the camera the entire time — direct, confident, slightly proud gaze.
+
+Camera: static locked-off frame (this is a selfie video call), 9:16 vertical, chest-up framing.
+
+Pace: conversational, matching the natural rhythm of the Azerbaijani audio.
+
+Language: Azerbaijani.
+```
+
+### Common Avatar mistakes
+
+- Writing Veo-style scene descriptions (Kling Avatar wants performance guidance, not scene-building — the IMAGE provides the scene)
+- Using generic emotional words ("happy", "sad", "angry") — use concrete physical descriptors
+- Heavy face occlusion in input image (hand covering mouth = bad lip-sync)
+- Noisy audio (background music in voice file = lip-sync drift)
+- Pacing instructions that contradict the audio's actual rhythm
+- Forgetting to specify the language
+
+### Sources (verified 2026-06-15)
+
+- [kling.ai/quickstart/ai-lip-sync-guide](https://kling.ai/quickstart/ai-lip-sync-guide)
+- [Higgsfield — Kling AI Avatar](https://higgsfield.ai/blog/Kling-AI-Avatar)
+- [Scenario — Kling Pro AI Avatar](https://help.scenario.com/en/articles/kling-pro-ai-avatar-the-essentials/)
+- [Pollo AI — Kling Lip Sync guide](https://pollo.ai/hub/how-to-use-kling-ai-lip-sync)
+
+## Image-to-Video
+
+### Mode A — Start frame only
+Provide first frame as reference image:
+```
+[Reference image: starting frame]
+Action: subject begins to walk forward, then stops
+Camera: static medium shot
+Duration: 5s
+```
+
+### Mode B — Start + End frame (key Kling 3.0 feature)
+Kling 3.0 has **native start-and-end-frame** support. Provide TWO reference images, model generates the transition. The most powerful continuity tool — director-decided per shot (CLAUDE.md "Video prompt qaydaları" #3).
+
+**Use Mode B when:**
+- Emotional arc (exhaustion → smile, neutral → reaction)
+- Specific action endpoint (hover → tap, closed → open)
+- Pose evolution (standing → bent → straight)
+- Continuity bridge between sequential clips
+
+**Kling UI / API:**
+- Upload `start_frame` (first image)
+- Upload `end_frame` (last image)
+- Prompt describes the transition
+
+**Prompt format:**
+```
+Start frame: [reference image — initial state]
+End frame: [reference image — final state]
+Action: gradual transition from start state to end state — describe HOW (timing, micro-motions, expression evolution)
+Camera: [shot type, angle, movement during transition]
+Setting: [environment — constant across both frames]
+Lighting: [setup — constant]
+Style: [cinematic notes]
+Duration: Ns (typically 3-6s for clean transitions)
+```
+
+**Sample Mode B prompt:**
+```
+Start frame: woman bent forward, hands on knees, exhausted, no expression
+End frame: same woman, same pose, but with a small genuine smile, eyes lifting toward horizon
+Action: 0-1s: heavy recovery breath, no expression change. 1-2s: subtle softening at corners of mouth. 2-3s: smile emerges gradually, eyes lift. 3-4s: final state held.
+Camera: close-up, 85mm, hand-held respired, very slow push-in (10cm closer over duration)
+Setting: pre-dawn park trail, warming light, mist still hanging
+Lighting: warming pre-dawn ambient, faint horizon rim on cheekbone
+Style: cinematic Wieden+Kennedy documentary, photorealistic, authentic smile (NOT posed)
+Duration: 4s
+Aspect: 9:16
+```
+
+**Mode B best practices:**
+- **Start and end frames must share environment** — same location, similar lighting, same subject (different expression/pose OK)
+- **Describe transition as time-segments** — "0-1s: ..., 1-2s: ..."
+- **Subject identity must remain consistent** — Kling 3.0 handles this well if both frames are clearly the same person
+- **Camera move can happen during transition** — slow push-in, drift, etc., but keep simple
+
+## Aspect Ratios
+
+- 16:9, 9:16, 1:1
+- Set as parameter, not in prompt text
+
+## Common Mistakes
+
+- Verbose natural language prose
+- Including audio cues (Kling doesn't generate audio)
+- Multiple camera movements
+- Multiple actions in sequence
+
+## Pricing tier
+$$$ — moderate-high
+
+## Speed tier
+Slow — 1-3 minutes per clip

package/templates/skills/video-prompt-engineer/knowledge/models/luma.md

@@ -0,0 +1,97 @@
+# Luma Dream Machine — Video Model Knowledge
+
+## Provider
+Luma AI
+
+## Type
+Diffusion (3D-aware)
+
+## Strengths
+
+- **Image-to-video** is strongest in this category
+- Smooth, fluid motion
+- Reasonable price
+- Fast generation
+- 3D scene understanding (some)
+- Good for character animation from still
+
+## Weaknesses
+
+- 5-second fixed duration
+- Less control over motion details
+- Output style somewhat homogenous
+- Limited resolution
+
+## Best Practices
+
+1. **Always use image-to-video** mode
+2. Provide highest quality starting image
+3. Brief motion description (Luma fills in)
+4. Don't over-specify
+5. Trust Luma's defaults for small details
+6. Single subject works best
+
+## Prompt Anatomy (image-to-video)
+
+```
+[Reference image upload]
+[Brief motion description]
+[Optional camera note]
+```
+
+## Sample Prompts
+
+### Character animation
+```
+[Image: portrait of man by window]
+The man slowly turns his head from the window to look toward the
+entrance. Subtle breathing. Camera static.
+```
+
+### Establishing shot animation
+```
+[Image: establishing shot of diner exterior]
+Slow push-in toward the entrance. Rain continues to fall. The neon
+sign flickers gently.
+```
+
+### Product shot
+```
+[Image: coffee cup on table]
+Steam rises from the cup. Camera holds completely still. The light
+shifts subtly as if a cloud passes outside.
+```
+
+## Text-to-Video
+
+Less reliable but possible:
+```
+A coffee cup on a wooden table, steam rising slowly. Soft window
+light. Static camera. 5 seconds.
+```
+
+## Best Use Cases
+
+- Animating still photographs
+- Bringing AI-generated images to life
+- Quick concept videos
+- Subtle motion additions
+- Character "breath" shots (subject barely moves)
+
+## Aspect Ratios
+
+- 16:9, 9:16, 1:1
+- Supports vertical for social
+
+## Common Mistakes
+
+- Expecting Veo-level detail
+- Asking for complex multi-action sequences
+- Long prompts (Luma works best with brief)
+- Forgetting to upload reference image
+
+## Pricing tier
+$$ — affordable
+
+## Speed tier
+Fast — 1-2 minutes

package/templates/skills/video-prompt-engineer/knowledge/models/pika-2.md

@@ -0,0 +1,76 @@
+# Pika 2.x — Video Model Knowledge
+
+## Provider
+Pika Labs
+
+## Type
+Diffusion
+
+## Strengths
+
+- Fast generation
+- Stylization options
+- Creative visuals
+- Affordable
+- Image-to-video supported
+- Sound effects in latest version
+
+## Weaknesses
+
+- Less photorealistic than Veo or Kling
+- Animation-leaning aesthetic
+- Short duration limit (3-4s)
+- Less control over specific details
+- Faces less consistent
+
+## Best Practices
+
+1. Lean into stylized/creative output
+2. Image-to-video for character consistency
+3. Short prompts work best
+4. Specify aspect ratio
+5. Use for B-roll, transitions, atmosphere
+6. Don't expect Hollywood realism
+
+## Prompt Anatomy
+
+```
+[Subject + style description]
+[Brief action]
+[Camera note]
+Duration: Ns
+```
+
+## Sample Prompt
+
+```
+A man in charcoal coat standing by rainy window in 1960s diner,
+melancholic mood, slowly turning head, cinematic shallow depth of
+field, warm tungsten lighting. Static medium shot. Duration: 4s.
+```
+
+## Best Use Cases
+
+- B-roll filler
+- Atmospheric transitions
+- Stylized music video clips
+- Creative experimental shots
+- Quick concept tests
+- Animated/illustrated styles
+
+## Aspect Ratios
+
+- 16:9, 9:16, 1:1, 4:3, 5:2
+
+## Common Mistakes
+
+- Expecting photorealistic detail
+- Long action sequences
+- Multiple characters in detail
+- Reading Pika as Veo replacement
+
+## Pricing tier
+$$ — affordable
+
+## Speed tier
+Fast — 30-60s

package/templates/skills/video-prompt-engineer/knowledge/models/runway-gen-4.md

@@ -0,0 +1,102 @@
+# Runway Gen-4 — Video Model Knowledge
+
+## Provider
+Runway
+
+## Type
+Diffusion
+
+## Strengths
+
+- Strong text-to-video and image-to-video
+- Multi-shot consistency improving
+- Large user community + tutorials
+- Director/style references work
+- Up to 10s clips
+- Stylization control
+- Camera move presets
+
+## Weaknesses
+
+- Pricier than alternatives
+- Veo more cinematic in head-to-head
+- Motion can feel "cartoony" in fast scenes
+- Less reliable for photorealistic faces
+
+## Best Practices
+
+1. Aesthetic-driven descriptions
+2. Director/photographer references work
+3. Image-to-video stronger than text-to-video
+4. Camera move keywords from preset list
+5. Aspect ratio explicit
+6. Style references via `--style`
+
+## Prompt Anatomy
+
+```
+[Cinematic style declaration with reference]
+[Subject + setting + action]
+[Camera move preset or "static"]
+[Mood / aesthetic / color grading]
+Duration: Ns
+```
+
+## Sample Prompt
+
+```
+Cinematic medium shot in the style of Wong Kar-wai, 1960s diner
+interior at night, melancholic mood. A man in a charcoal coat stands
+by a rain-streaked window, slowly turning his head to look toward
+the entrance. Camera: static medium shot, eye-level. Lighting: warm
+tungsten key with cool blue window backlight. Color grading: muted
+warm tones with cool shadows, slight film grain. Shallow depth of
+field. Duration: 6s.
+```
+
+## Image-to-Video Strength
+
+Runway excels at animating still images:
+1. Generate image elsewhere (Nano Banana, Midjourney)
+2. Upload to Runway
+3. Add motion description
+4. Generate video
+
+```
+[Reference image: portrait of subject]
+Animation: subject begins to slowly turn head left, eyes following.
+Subtle breathing motion. Hair slightly catches air movement. Static
+camera. Duration: 5s.
+```
+
+## Camera Presets
+
+Runway has built-in camera presets:
+- Static
+- Pan left/right
+- Tilt up/down
+- Dolly in/out
+- Zoom in/out
+- Orbit left/right
+- Crane up/down
+- Tracking left/right
+
+Specify by name in prompt.
+
+## Aspect Ratios
+
+- 16:9, 9:16, 1:1, 21:9
+- Wide format support
+
+## Common Mistakes
+
+- Treating like Veo (different model, different prompt style)
+- Forgetting style references
+- Combining too many camera movements
+- Long verbose prompts (medium length is sweet spot)
+
+## Pricing tier
+$$$$ — premium
+
+## Speed tier
+Fast — 30-90s

package/templates/skills/video-prompt-engineer/knowledge/models/seedance-2.md

@@ -0,0 +1,80 @@
+# Seedance 2 — Video Model Knowledge
+
+## Provider
+ByteDance (China)
+
+## Type
+Diffusion
+
+## Strengths
+
+- **Stylized aesthetic** — fashion, beauty, social media-friendly
+- Smooth motion
+- TikTok-friendly outputs (vertical, dynamic)
+- Trend-aware aesthetic
+- Free tier available
+- Fast generation
+
+## Weaknesses
+
+- Western/cinema look weaker than Veo
+- Limited audio support
+- Less photorealistic for serious drama
+- Documentation primarily Chinese
+- Less control than Veo or Kling
+
+## Best Practices
+
+1. Lean into stylization — don't fight it
+2. Subject + visual aesthetic descriptors
+3. TikTok/Reel framing works best
+4. Vertical 9:16 native
+5. Short durations (4-5s optimal)
+6. Trend-aware references work
+
+## Prompt Anatomy
+
+```
+[Subject + visual style description]
+[Action and camera direction]
+[Aesthetic notes — fashion, beauty, trend]
+Duration: Ns
+```
+
+## Sample Prompt
+
+```
+A young woman with long flowing dark hair, wearing a white silk dress,
+walks through a softly-lit forest at golden hour. Camera tracks
+alongside her in slow motion. Sunlight filters through leaves creating
+dappled light on her face and dress. Soft pastel color grading,
+romantic dreamy aesthetic, fashion editorial style, vertical 9:16
+format. Duration: 5s.
+```
+
+## Best Use Cases
+
+- Fashion content
+- Beauty product reels
+- Lifestyle TikToks
+- Music video B-roll (stylized)
+- Aesthetic mood reels
+
+## Aspect Ratios
+
+- 9:16 vertical (primary)
+- 16:9 supported
+- 1:1 supported
+
+## Common Mistakes
+
+- Trying to force serious drama (use Veo instead)
+- Long prompts with technical detail
+- Unrealistic Western cinematic references
+- Detailed dialogue scenes
+
+## Pricing tier
+Free / $$ — accessible
+
+## Speed tier
+Fast — 30-90s