@azerogluemin/ai-bootstrap 0.2.1

package/templates/skills/video-prompt-engineer/knowledge/camera-movements.md

@@ -0,0 +1,192 @@
+# Camera Movements — Video Generation
+
+> How cameras move in cinema. Critical for video prompts because each model handles movement differently.
+
+## Movement types
+
+### Static (default — RECOMMENDED for AI video)
+- Camera doesn't move
+- Subject can move within frame
+- Most stable, lowest artifact risk
+- Best for dialogue, reaction shots, contemplative moments
+- All AI models excel at static
+
+### Pan (horizontal rotation)
+- Camera rotates left or right on axis
+- Subject typically tracks across
+- Specify: "slow pan left 90°" or "quick whip pan right"
+- Works in: Veo, Runway, Sora; risky in Kling, Pika
+
+### Tilt (vertical rotation)
+- Camera rotates up or down on axis
+- Reveals subjects taller than frame
+- Specify: "slow tilt up to reveal building"
+- Works similarly to pan
+
+### Dolly (forward/backward)
+- Whole camera moves toward or away
+- "Dolly in" creates intimacy/tension
+- "Dolly out" reveals context, creates distance
+- Specify: "slow dolly in 2 feet"
+- Works in Veo, Sora; harder in Kling
+
+### Truck / Track (sideways)
+- Camera moves left or right
+- Like pan but whole camera shifts
+- Used for following subject walking
+- Specify: "tracking shot, camera moves right with subject"
+
+### Crane / Jib (vertical move)
+- Camera moves up or down
+- Can combine with tilt
+- Reveals scale, drama
+- Specify: "crane up from ground level to overhead"
+
+### Handheld
+- Subtle natural shake
+- Documentary feel, intimacy
+- Specify: "subtle handheld camera, slight movement"
+
+### Steadicam / Gimbal
+- Smooth glide as if floating
+- Long takes following subject
+- Specify: "smooth steadicam following subject through hallway"
+
+### Zoom (lens, not camera)
+- Camera doesn't move; lens focal length changes
+- "Zoom in" / "Zoom out"
+- Less cinematic than dolly
+- Use sparingly
+
+### Push-in / Pull-out (combination)
+- Dolly + zoom hybrid
+- "Vertigo effect" / Dolly zoom: dolly forward + zoom out (or opposite)
+- Iconic Hitchcock effect
+- Very hard for AI
+
+## CRITICAL RULE for AI video
+
+**Combine movements with caution.**
+
+- Single movement: ✅ usually works
+- Two simultaneous: ⚠️ risky in most models
+- Three+: ❌ fails or produces artifacts
+
+Veo handles single movements best. Kling prefers static. Sora attempts combinations.
+
+## Movement speed
+
+- Slow — most cinematic, most stable for AI
+- Medium — natural
+- Fast / whip — high artifact risk, use only when style demands
+
+## Prompt vocabulary
+
+GOOD:
+```
+Camera: static medium shot, eye-level, no movement.
+```
+
+GOOD:
+```
+Camera: slow dolly in from medium to close-up over the duration,
+following subject's eyes. No pan or tilt.
+```
+
+BAD:
+```
+Dynamic camera with cool moves.
+```
+
+BAD:
+```
+Camera pans left while dollying in and tilting up.  // Too much
+```
+
+## Movement-to-emotion mapping
+
+- Static → contemplation, dialogue, focus
+- Slow dolly in → intimacy, tension building
+- Slow dolly out → reveal, isolation, ending
+- Tracking → following journey, momentum
+- Handheld → urgency, realism, chaos
+- Crane up → freedom, scale, finality
+- Crane down → entering scene, intimacy
+
+## Default for AI video
+
+When unsure: **static**. Add movement only when scene demands it.
+
+---
+
+## HARD RULE — Camera motion must be smooth. No shake, no wobble, no jitter (default)
+
+**By default, every camera movement in a video prompt must be SMOOTH and MECHANICALLY STABLE** — as if mounted on a polished dolly track, a motion-controlled crane, or a high-end gimbal. Zero handheld float, zero micro-jitter, zero wobble, zero shake, zero parallax tremor.
+
+### Why this rule is non-negotiable
+
+Image-to-video models (Kling, Runway, Hailuo, Sora, Veo) have a strong **default bias toward adding handheld instability** — even when the prompt asks for a static or smooth move. They interpret "natural" as "amateur handheld" because most of their training data is consumer footage. Without an explicit anti-shake instruction, the output will almost always carry visible micro-jitter that ruins the cinematic feel and creates motion-sickness in long sequences.
+
+This is a **default rule** — applies to every shot unless the script explicitly demands handheld documentary aesthetic.
+
+### How to enforce smoothness in the prompt
+
+Add an explicit smoothness clause to **every camera description**, regardless of movement type:
+
+GOOD — explicit smoothness:
+```
+Camera: one single continuous slow dolly push-in along the street axis.
+Movement is mechanical-smooth, as if on a polished dolly track or
+motion-controlled crane. Zero handheld float, zero micro-jitter, zero
+wobble, zero shake, zero camera-body tremor. No rotation, no tilt, no
+parallax sliding. Perfectly stabilized motion throughout.
+```
+
+GOOD — explicit smoothness for static:
+```
+Camera: completely static and locked off. Tripod-stable. Zero drift,
+zero breathing, zero handheld float, zero micro-jitter. Frame edges
+do not move. Treat as a locked tripod shot.
+```
+
+BAD — silent on smoothness:
+```
+Camera: slow dolly in.    // Model will add wobble by default
+```
+
+BAD — silent on smoothness for "static":
+```
+Camera: static medium shot.    // Model will still drift
+```
+
+### Negative-prompt block for shake (paste into negative-prompt field when available)
+
+```
+camera shake, handheld wobble, micro-jitter, frame jitter, parallax
+sliding, drifting frame, breathing camera, vlogger handheld, amateur
+shake, motion-sickness wobble, unstable frame, drifting horizon, tilted
+frame drift
+```
+
+### When handheld IS the intent (exception)
+
+Only when the **script explicitly calls for documentary handheld** (Saving Private Ryan / Children of Men aesthetic, urgent vérité scene, found-footage horror, etc.) do you allow shake — and even then, specify the type and amount:
+
+```
+Camera: subtle controlled handheld, gentle organic breathing motion
+(2–3% drift), no jerky moves, no whip pans. War-documentary aesthetic
+in the tradition of Janusz Kamiński's Saving Private Ryan operator
+work — alive, but composed.
+```
+
+Notice: even "intentional handheld" is *controlled* — never the raw amateur jitter the model produces by default.
+
+### The two-line smoothness block (drop into every prompt)
+
+Until this rule becomes muscle memory, append this exact block to the Camera section of every prompt:
+
+```
+Movement is mechanical-smooth — as if on a polished dolly track or motion-controlled gimbal. Zero handheld float, zero micro-jitter, zero wobble, zero shake. Perfectly stabilized cinematic motion.
+```
+
+This single addition consistently eliminates the most common "AI video looks amateur" failure mode.

package/templates/skills/video-prompt-engineer/knowledge/continuity-rules.md

@@ -0,0 +1,152 @@
+# Continuity Rules — Video Generation
+
+> When making multiple clips that should connect (a scene of multiple shots), continuity must be maintained explicitly.
+
+## What continuity means
+
+Across multiple clips that depict the same scene:
+- Same character looks identical
+- Same lighting setup
+- Same wardrobe (unless intentional change)
+- Same props in same positions
+- Time of day consistent
+- Same general environment
+
+## The 4 continuity dimensions
+
+### 1. Spatial continuity
+Position of subjects and objects in space.
+- If subject ended on right side of frame, next shot should start with them on right
+- Eyeline match: where subject was looking, next shot should reflect
+
+### 2. Temporal continuity
+Time within scene.
+- If it was raining, next clip rains (unless time-passed cut)
+- If subject was holding cup, they should still hold it (or visibly put it down)
+
+### 3. Visual continuity
+Look and style.
+- Same lighting setup
+- Same color grading
+- Same lens character
+
+### 4. Audio continuity
+- Same ambient sound bed
+- Music tempo/mood consistent
+- Voice quality matches
+
+## Continuity prompt strategy
+
+### For each clip after the first
+
+Include explicit continuity reference:
+
+```
+CONTINUITY FROM PREVIOUS CLIP:
+- Subject: [same character description]
+- Wardrobe: [exact same outfit]
+- Lighting: [exact same setup]
+- Props: [what they're holding, what's around them]
+- Position: [where they are in space]
+- Mood: [emotional state]
+
+THIS CLIP'S NEW ACTION:
+[describe what changes]
+```
+
+### Example: 3-clip scene
+
+#### Clip 1 — Establishing
+```
+[Wide shot]
+A 35-year-old man enters a 1960s diner from the rain. He removes his
+coat and hangs it. Walks to counter. Sits on stool.
+
+[Audio: rain, door bell, footsteps, jazz from jukebox]
+Duration: 8 seconds.
+```
+
+#### Clip 2 — Medium (continuity)
+```
+CONTINUITY: Same man, now seated at counter from clip 1. Wet hair
+from rain, coat now hanging on hook visible in background. Same
+warm tungsten + cool window lighting. Jazz still playing.
+
+[Medium shot, eye-level]
+He looks down at the menu, then up at the empty stool beside him.
+
+[Audio: jazz continues, his quiet breathing, distant rain]
+Duration: 6 seconds.
+```
+
+#### Clip 3 — Close-up (continuity)
+```
+CONTINUITY: Same man from clips 1 and 2. Wet hair, looking at empty
+stool. Same lighting. Same audio bed.
+
+[Close-up, slight low angle]
+His expression shifts from neutral to melancholic. A subtle exhale.
+He looks back down at the menu.
+
+[Audio: jazz, his exhale]
+Duration: 5 seconds.
+```
+
+## Reference image strategy
+
+For best continuity:
+1. Generate Clip 1
+2. Extract last frame
+3. Use as reference image for Clip 2
+4. Repeat
+
+Models that support this well:
+- Runway (image-to-video)
+- Luma (image-to-video)
+- Kling (with first frame reference)
+
+## Common continuity errors
+
+### Wardrobe drift
+Solution: re-state outfit in EVERY clip prompt.
+
+### Lighting flicker
+Solution: lock lighting setup in every prompt with same exact words.
+
+### Hair changes
+Solution: re-state hair description identically.
+
+### Background props move
+Solution: simplify background, re-list in continuity.
+
+### Time-of-day shift
+Solution: explicitly state time of day each time.
+
+## Match cuts
+
+When you want a deliberate visual match between clips:
+- Same composition
+- Same focal point
+- Match action mid-motion
+
+```
+Clip 1 ends: subject reaches for cup
+Clip 2 begins: hand grasping cup at exactly same position
+```
+
+This requires very precise prompting.
+
+## Eyeline match
+
+When two characters look at each other:
+- Subject A looks screen-right at 30° down
+- Subject B should be looking screen-left at 30° up
+- Heights must match
+
+## Default approach
+
+For multi-clip scenes:
+1. Plan all clips upfront
+2. Write continuity-aware prompts
+3. Re-state EVERYTHING that should stay constant
+4. Only describe what's changing

package/templates/skills/video-prompt-engineer/knowledge/duration-pacing.md

@@ -0,0 +1,130 @@
+# Duration & Pacing — Video Generation
+
+> How long should AI video clips be, and how to pace action within that duration.
+
+## Duration capabilities by model
+
+| Model | Min | Max | Optimal |
+|-------|-----|-----|---------|
+| Veo 3.1 | 4s | 8s | 6-8s |
+| Kling 2.x | 5s | 10s | 5-7s |
+| Seedance 2 | 4s | 6s | 4-5s |
+| Runway Gen-4 | 5s | 10s | 5-7s |
+| Sora | 4s | 60s+ | 5-15s |
+| Pika 2 | 3s | 4s | 3-4s |
+| Luma | 5s | 5s | 5s (fixed) |
+| Hailuo | 4s | 6s | 4-6s |
+
+## Optimal duration by use case
+
+### Reaction shot (subtle): 4-5s
+- Single emotion beat
+- Facial expression change
+- Static camera
+- Minimal motion
+
+### Establishing shot: 5-6s
+- Slow reveal
+- Subject in environment
+- Optional slow camera movement
+- Atmospheric
+
+### Action beat: 6-8s
+- One complete action
+- Walk to door, sit down, raise glass
+- Camera follows or static
+- Clear start/end
+
+### Dialogue clip: 6-8s
+- 2-3 sentences max
+- Lip sync visible
+- Static or slow camera
+- Audio cue for delivery
+
+### Transition / B-roll: 4-5s
+- Quick visual moment
+- Mood/atmosphere
+- Static or single move
+
+## Pacing within clip
+
+### 4-second clip
+```
+0.0s — Start state established
+0.5-3.0s — Single action beat
+3.0-4.0s — End state held
+```
+
+### 6-second clip
+```
+0.0-0.5s — Hold start state
+0.5-2.0s — Action begins
+2.0-4.0s — Action continues
+4.0-5.5s — Action completes
+5.5-6.0s — End state held
+```
+
+### 8-second clip
+```
+0.0-1.0s — Establishing beat (subject visible, scene set)
+1.0-3.0s — Action initiation
+3.0-5.0s — Action peak
+5.0-7.0s — Action resolution
+7.0-8.0s — End state, hold
+```
+
+## Prompt phrases for pacing
+
+```
+Duration: 6 seconds.
+The action unfolds slowly:
+- Frames 0-1: subject is still, looking forward
+- Frames 1-4: subject begins to turn head left, gradually
+- Frames 4-5: head reaches final position
+- Frames 5-6: subject holds, eyes finishing the look
+```
+
+This timing breakdown helps Veo and Sora especially.
+
+## Common mistakes
+
+### Too much in too little time
+```
+BAD: In 4 seconds, the subject walks across the room, sits down,
+picks up a cup, takes a sip, looks out the window, and sighs.
+```
+Result: rushed, jittery, artifacts.
+
+### Too little in too much time
+```
+BAD: 10-second clip of someone blinking once.
+```
+Result: model fills with random motion.
+
+### No duration specified
+```
+BAD: A man turns his head.
+```
+Result: model picks duration randomly, often poorly.
+
+## Format-specific durations
+
+### Social media reel
+- 4-6s clips that loop or chain together
+- Multiple short clips edited externally
+
+### Cinematic short film
+- 5-8s clips
+- Plan for editing transitions
+
+### Brand commercial
+- 6-8s hero clips
+- Multiple beats per scene
+
+### Music video
+- 4-6s clips
+- Cut to beat externally
+
+## Default for AI video
+
+When unsure: **6 seconds**. Long enough for one beat, short enough to stay stable.

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

@@ -0,0 +1,78 @@
+# Hailuo / MiniMax Video — Video Model Knowledge
+
+## Provider
+MiniMax (China)
+
+## Type
+Diffusion
+
+## Strengths
+
+- **Free tier** available
+- Decent text-to-video
+- Asian features render well
+- Up to 6-second clips
+- Image-to-video supported
+- Active development
+
+## Weaknesses
+
+- Limited English documentation
+- Variable quality (some hit-or-miss)
+- Less control than premium models
+- Less reliable for complex scenes
+
+## Best Practices
+
+1. Keep prompts concise
+2. Single clear subject and action
+3. Static or minimal camera movement
+4. Image-to-video for better control
+5. Test multiple times for best take
+
+## Prompt Anatomy
+
+```
+[Subject description]
+[Single action]
+[Camera note]
+[Style note]
+Duration: Ns
+```
+
+## Sample Prompt
+
+```
+A young man in dark coat standing by a rainy window, slowly turning
+his head to look toward the door. Camera static medium shot. Warm
+interior lighting. Cinematic style. Duration: 5s.
+```
+
+## Best Use Cases
+
+- Free / low-budget projects
+- Concept testing
+- Social media B-roll
+- Asian content (where features render well)
+- Quick experimentation
+
+## Aspect Ratios
+
+- 16:9, 9:16, 1:1
+
+## Common Mistakes
+
+- Long verbose prompts
+- Multiple complex actions
+- Expecting Hollywood quality on free tier
+- Not iterating (quality variable)
+
+## Pricing tier
+Free tier / $ paid
+
+## Speed tier
+Medium — 1-3 minutes
+
+## Summary
+
+Hailuo is best when budget is constrained. For premium output, use Veo or Kling. For experimentation, Hailuo's free tier is valuable.