git:20260903.1675451 to git:20260906.c2f2b15

45 added, 96 removed. Audit A to A.

---

name: drone-fly-on-city-fireworks

- description: Create cinematic photoreal MiniMax H3 night-city drone flights with a route-controlled 360-degree landmark orbit and physically continuous fireworks, smoke, reflections, exposure and native sound. Use when the user wants an aerial skyline celebration with loaded scene and route references; not for daylight flights, static images or character-led scenes.

+ description: "Direct a three-phase MiniMax H3 fireworks FPV flight: launch near ground in P1, complete one translated 360-degree orbit around P1's primary scene subject, then follow verified P2 bends to its endpoint; use a collision-safe FPV scene fallback when P2 is unreadable. Generate P3 onward every five seconds from P1 pixels without assuming a city."

---

- # Drone City Fireworks Director for MiniMax H3

- Apply this Special Skill with the bound Default H3 Prompt Writing Skill. Produce an editable, chronological Director Design whose camera flight and fireworks can be compiled directly into H3. Preserve the user's exact duration, aspect ratio, named city, landmark, weather, start/end composition and MUSIC setting.

- ## 1. Required Reference Roles

- Require two genuinely loaded Media Pool Pictures:

- - `@P1` is the scene master and opening frame. Preserve its real skyline geometry, landmark count and spacing, roads, weather, time, light direction, exposure, colour grade, atmosphere, lens height and visual effects.

- - `@P2` is route-only analysis data. Extract only its start, meaningful turns, curve direction and endpoint, then express those as abstract camera movement. It must never become an H3/Z-Image visual reference, Timeline clip, start/end frame, style source, composition source or identity anchor.

- If either reference is missing, hard-block and name the missing Picture. Never invent a Media Pool ID.

- For Design JSON, explicitly register P1 as `media_type="image"`, `usage="h3_reference"`; register P2 as `media_type="image"`, `usage="analysis_only"`, `reuse_policy="whole_design"`. P2 may remain in the Virtual Media Pool but consumes no physical H3 Picture slot and is absent from all uploads and loaders.

- ## 2. Route and Orbit

- Reduce P2 into 3–6 meaningful planning waypoints in normalized screen coordinates. Preserve every important bend and the original start/end direction; never claim GPS, metres or real telemetry from a 2D drawing. Before rendering, translate waypoint data into natural physical camera language without P2, WP labels, coordinates, red lines, map graphics or navigation terminology.

- Default to one smooth wide clockwise orbit around the named landmark while gradually rising. Unless the user asks otherwise:

- - begin low and forward-left, facing the landmark;

- - advance with stable altitude transition and realistic drone inertia;

- - complete exactly one continuous 360-degree orbital yaw;

- - keep the target readable and the horizon level;

- - use natural parallax and exposure adaptation;

- - avoid towers, roofs, cables, trees, traffic and people;

- - finish the moving orbit high on the right, let the active sequence finish, then return to P1's exact original camera position and composition for the final one-second plate hold.

+ # P1-Locked Drone Fireworks Director

- Never reverse direction, repeat the orbit, teleport, pass through architecture or allow the skyline geometry to flicker.

+ Use with the Default H3 Prompt Writing Skill. P1 determines what exists; P2 determines how the camera travels; the fireworks profile determines only celebration effects. A location or landmark named in an example must never override P1.

- ## 3. Shot and Segment Plan

+ ## 1. Reference Roles

- Use chronological, gap-free, non-overlapping Shot Blocks on the 0.5-second grid. Use roughly one Shot per 4–6 seconds and one primary camera movement per Shot. Each Shot must state:

+ - `@P1` is the sole scene master and opening composition. Keep it active in every Segment. Preserve the actual location, landmark identity/count/spacing, architecture, roads, objects, weather, time, lighting, colour, exposure, atmosphere, horizon, camera height and lens character visible in P1.

+ - `@P2` is local route-analysis data only. Extract its red stroke's start, bends, direction and endpoint. Never put P2 in the Timeline, upload list, H3/Z-Image Loader, start/end frame, style, composition or identity reference.

+ - Hard-block a genuinely missing P1 or P2; never invent a label.

- - must-complete camera/subject action;

- - environment response;

- - incoming continuity state;

- - outgoing continuity state;

- - optional flourish that may be dropped before any required action.

+ Register P1 as full-duration `h3_reference/whole_design` and P2 as full-duration `analysis_only/whole_design`.

- For 15 seconds, prefer three phases:

+ P1 also determines the sky and time of day. Do not turn a daylight, sunset or warm-coloured P1 into a cool-blue night scene merely because fireworks were requested. Keep the base palette and exposure; allow only temporary illumination from visible bursts. An authored time-of-day change must be explicit, not inherited from a Skill example.

- 1. **Approach / ignition**: establish the landmark and begin the orbit; only a few distant rooftop launches appear.

- 2. **Side reveal / escalation**: gold, white and deep-red bursts expand behind and above the towers; glass and wet streets respond.

- 3. **High-right finale / P1 return**: the moving orbit resolves and its fireworks complete; earlier smoke disperses while the camera returns to P1's exact original composition. The final second directly displays the immutable P1 plate with a new frozen celebration-effect layer.

+ ## 2. Ground Launch, P1 Orbit, Then P2 Route

- Scale the phases proportionally for other durations. Do not make every Shot a new location or reset the fireworks between cuts.

+ Studio locally extracts the largest continuous red stroke in P2 and reduces it to ordered path points. Every Shot must show forward translation through those bends with realistic inertia, parallax and a stable horizon. Route-follow is the default. An open S-curve, zig-zag or point-to-point path must never become an in-place rotation or generic circle.

- ## 4. Fireworks Physics and Continuity Ledger

+ Execute the movement in three non-overlapping phases: launch from a physically safe near-ground point in the P1 scene; physically fly one complete, smooth, wide clockwise lap around P1's primary subject at a constant safe radius, translating in order past the front/start side, right side, rear, left side and back near the front/start side while the rigid FPV camera stays aligned with the drone nose and instantaneous forward tangent. It must never independently yaw, pan or gimbal-lock toward the subject. Let the landmark travel naturally along the inside edge, pass behind the camera when geometry requires it, and reappear as the aircraft advances; use strong background parallax to prove translation. Only after the lap, enter the verified route at its green start, follow every bend, and reach its blue endpoint. This is never an in-place camera rotation, continuous subject-centred look-at, panoramic yaw, barrel roll, optical spin or rotating background. Fireworks timing must not override this order.

- Fireworks are visible environmental events, never graphic overlays. For every Shot record and preserve:

+ For the 15-second template, allocate 0–2 seconds to launch, 2–10 seconds to the complete physical lap and only 10–15 seconds to the route. Lock all route-exit motion until the aircraft has visibly returned near the lap's front starting side.

- - launch zone and burst position relative to the landmark;

- - colour and burst family;

- - ignition, radial expansion, spark falloff and extinction phase;

- - smoke amount, wind direction and carried position;

- - reflection state on glass façades, wet roofs and streets;

- - exposure response and return to the cool-blue night baseline;

- - last audible boom/crackle state at the outgoing boundary.

+ Use a fully immersive first-person FPV camera rigidly attached to the aircraft, with mild GoPro-like ultra-wide fisheye, speed-driven motion blur, strong inertia and physically motivated banking. During the P1 lap, use moderate coordinated banking only and keep the horizon readable. Dives, inversion and up to 180-degree rolls are permitted only after the full lap is complete and the verified route or fallback phase has begun. Keep architecture, fireworks and trees geometrically stable through every permitted roll; no frame tearing, slow mechanical god view or ordinary camera push.

- Each shell follows physical order: a distant launch source or rising spark, aerial burst, expanding particles, delayed boom according to distance, falling embers, smoke drift and fading reflections. Bursts occur behind or above the landmark and must not touch, wrap around, emerge from, replace or deform the towers. Avoid a continuous wall of fireworks: keep skyline depth, negative space and readable architecture.

+ If P2 is unverified, report `ROUTE NEEDS REVIEW`, retain the launch and 360-degree P1 orbit, and replace only the route phase with a collision-safe FPV scene flight: low ground skim; hard left bank and climbing arc; brief inverted crest into a dive; right counter-roll around visible obstacles; tight figure-eight crossover; level sprint toward a safe opening or distant horizon. Do not invent a P2 route endpoint.

- The final plate's gold/white/deep-red bursts are discrete radial particle blossoms behind and above the spires. They are not orbit rings, halos, solid discs, neon loops, energy circles or light ribbons. Smoke in the moving sequence drifts consistently; the earlier moving smoke may fully disperse during the return, after which the final plate receives only its own thin frozen smoke layer.

+ ## 3. P1-Derived Five-Second Scene References

- ## 5. Scene Keyframe Chain and Immutable P1 Effects Plate

+ Create P3 as the frozen 0–5 second opening scene anchor using the lowest denoise, then create P4 onward at exact five-second boundaries. A 35-second design uses P3–P9; a 45-second design continues through P11. Build every Picture using actual P1 image pixels via `source_plate_mode=p1_img2img` and `source_plate_media_id=P1`. BLIP/AI descriptions are supplementary only. Every Picture must preserve P1's primary subject building, place identity, landmark count, architecture, composition, surrounding scene, roads, sky, weather, time, colour palette, colour temperature, light, exposure, atmosphere and lens.

- With only P1/P2, use P1 for the moving scene and create one terminal P1 effects plate. If the user has loaded and selected P3/P4/P5 or later city-scene Pictures, treat them as an ordered user-authored chain while skipping analysis-only P2:

+ P1 remains active for the whole video; each five-second interval may add only its current scene reference. P3 and P4 onward are supporting scene states of P1, not new scene masters. P3 contains no drawn motion; the 360-degree orbit exists only in H3 Shot motion. P2 is never their visual parent.

- - each scene Picture owns one disjoint `time_scoped` interval;

- - no future Picture is loaded while an earlier interval renders;

- - each ownership interval is a separate native H3 Job;

- - only the preceding final 24 silent video frames cross a boundary as motion context;

- - no previous audio crosses the boundary;

- - create exactly one automatic environment-only terminal keyframe after the latest user anchor;

- - leave its Picture ID unassigned so Studio chooses the next genuinely empty P slot.

+ At execution time, each P1-derived stage exclusively replaces the direct P1 loader for its interval. Split rendering at every five-second boundary and send only the current stage Picture to H3; never send P1, P3, P4 and P5 together. All stage Pictures describe the same single scene instance over time. Render one instance of the primary landmark or landmark group only. If P1 contains a paired landmark, preserve the original pair exactly once; no second pair, mirrored copy, cloned building or repeated landmark is allowed.

- The terminal request must set `source_plate_media_id="P1"`, `source_plate_mode="immutable_effect_composite"`, `source_plate_effect_profile="fireworks"`, `immutable_scene_plate=true` and `final_hold_seconds=1.0`. It does not inherit or redesign the last scene. P1 is the immutable scene plate: preserve its exact camera position, altitude, focal length, framing, horizon height, landmark size/spacing, skyline geometry, roads, buildings, weather, cool-blue grade, baseline exposure, contrast and original city lights. Add only separate fireworks particles, thin smoke and permitted warm reflection/illumination effects on top. The final second is perfectly static.

+ Each stage is a frozen photograph, not a movement diagram. Remove orbit, 360-degree, yaw, route, trajectory and waypoint motion from still prompts and keywords. Append exactly:

- The moving fireworks and earlier smoke complete and disperse during the transition back to P1. The terminal plate then shows a newly composed frozen celebration moment; this effect layer is not motion history and cannot move or reshape P1 pixels. It never introduces a person or changes the city.

+ `Clean photographic scene with unobstructed architecture, natural sky and physically plausible lighting. Preserve the source image's scene, colour palette and exposure.`

- ## 6. Z-Image Still Isolation

+ Append this fireworks still contract:

- Every generated Picture is one frozen photographic moment, never a motion diagram. Remove 360-degree, orbit, orbital yaw, circle-path, trajectory, route and waypoint clauses from ordinary still-image prompt sentences and `subject_keywords`. Keep the city, architecture, weather, lighting, fireworks state, smoke, reflections, lens and one frozen camera position.

+ `Fireworks are separate radial particle bursts located behind and above the skyline, with individual sparks, natural smoke and physically plausible reflections. Keep architectural silhouettes clearly readable.`

- Append exactly:

+ Use the request-scoped negative prompt:

- `The drone flight path is implied only through camera motion and must never be visible in the image. No orbit ring, no circular light trail, no glowing ellipse, no trajectory line, no HUD, no graphic overlay around the towers.`

+ `visible flight path, orbit ring, circular light trail, glowing ellipse, light ribbon, trajectory line, energy ring, HUD overlay, graphic circle, neon loop around buildings, continuous firework ring around buildings, fireworks forming a flight path, fireworks wrapped around towers, solid neon fireworks, duplicated landmark, fused towers`

- Also append this fireworks clarification:

+ Never copy this negative list into an H3 video prompt.

- `Fireworks are separate radial particle bursts located behind and above the skyline, with individual sparks, natural smoke and physically plausible reflections; they never form a continuous ring, ribbon, ellipse or flight path around any building.`

+ ## 4. Fireworks Physics and Continuity Ledger

- Store this request-scoped Z-Image negative prompt:

+ For each Shot track burst location relative to existing P1 architecture, colour, launch/expansion/falloff/extinction phase, smoke amount and wind direction, surface reflections, exposure response and delayed boom/crackle state. Use the physical order: launch spark, aerial burst, discrete radial particles, distance-delayed boom, falling embers, drifting smoke and fading reflections.

- `visible flight path, orbit ring, circular light trail, glowing ellipse, light ribbon, trajectory line, energy ring, HUD overlay, graphic circle, neon loop around buildings, continuous firework ring around buildings, fireworks forming a flight path, fireworks wrapped around towers, solid neon fireworks, duplicated landmark, fused towers`

+ Fireworks remain behind or above existing scene geometry. They never touch, cover, wrap around, emerge from, replace or deform a building. Keep negative sky space and readable architecture; do not create a continuous ring or wall of effects.

- Never copy that negative list into an H3 video prompt. P2 is never a visual parent of any generated Picture.

+ ## 5. Native Continuous Ending

- ## 7. H3 Prompt Contract

+ Let the H3-generated camera reach the authored endpoint naturally. Keep fireworks, smoke and environmental motion continuous. No automatic terminal image, forced P1 return, one-second freeze or local fireworks composite is requested. Do not reserve an extra final Segment for a still image. Preserve the generated ending during output assembly.

- Compile each Segment into a single chronological English H3 prompt. Match visual description, camera movement, fireworks state and sound to the same time order. The prompt must:

+ ## 6. H3 Prompt and Native Sound

- 1. establish P1's city geometry, landmark, cool-blue night grade, realistic architecture and wide-angle aerial lens;

- 2. describe the continuous physical flight and one clockwise orbit;

- 3. describe distinct gold, white and deep-red particle bursts behind/above the skyline, smoke drift, glass/wet-street reflections and natural exposure adaptation;

- 4. resolve the moving sequence high-right, disperse its earlier smoke, return to P1's exact original framing, then display the immutable P1-plus-effects plate for one second;

- 5. end with positive clean-frame language: realistic unobstructed city imagery, non-visual planning controls kept off-screen, stable horizon, coherent city geometry and continuous aerial inertia.

+ Each H3 Segment prompt follows one chronology: P1 scene facts, ground-launch phase, proportional physical lap progress with front/right/rear/left/front position checkpoints, later verified route displacement or FPV fallback, fireworks state, environment response, continuity state and diegetic audio. Lap degrees apply only before the route begins. During the lap, keep the camera rigidly forward along the flight tangent, let the building cross the inside edge naturally, use background parallax to prove translation, and explicitly reject subject-centred look-at, independent yaw, camera spin, panoramic pan and barrel roll. Reach the route endpoint only when P2 is verified. Never name P2 or expose control graphics in renderable prose.

- Do not name P2, red routes, waypoint graphics or the Z-Image negative list in the final H3 prompt. Do not add people, text, subtitles, logos, watermarks, duplicate towers, cartoon styling, unexplained spotlights or extra landmarks.

+ Use continuous high-altitude wind, distant local ambience, launch hiss, distance-delayed low-frequency booms and short crackle tails. Do not restart or cut effects at Shot boundaries. Add no dialogue or narration unless the Timeline contains it. Music obeys `MUSIC: OFF / AUTO / TIMELINE`.

- ## 8. Native Audio and Music

+ ## 7. Apply Gate

- Treat sound as diegetic and distance-aware: continuous high-altitude wind, distant city traffic bed, launch hiss, delayed low-frequency firework booms, short crackle tails and reflections from nearby towers. Preserve ambience across Shot boundaries; do not restart or abruptly cut a boom, crackle or smoke-related event at a Segment edge. Do not add dialogue, narration or studio voice unless the Timeline explicitly contains it.

+ 1. P1 is active in every Segment and the first generated moment establishes a near-ground takeoff inside its world; no example location replaces it.

+ 2. P2 is analysis-only and absent from every visual Loader.

+ 3. H3 launches first, completes one translated 360-degree orbit around P1's primary building second, then follows every verified P2 bend to its endpoint; these phases are never simultaneous.

+ 4. P3 is the 0–5 second opening anchor; P4 onward are five-second P1-derived scene states with no new scene identity.

+ 5. Current intervals use P1 plus only the relevant stage Pictures; future stages do not leak backward.

+ 6. Fireworks, smoke, reflection, exposure and audio states remain physically continuous.

- Music follows the homepage `MUSIC: OFF / AUTO / TIMELINE` mode. `OFF` means no score. `AUTO` may use restrained cinematic celebratory music that ducks beneath major firework impacts. `TIMELINE` uses only the authored music instruction.

+ ## Supported P2 Drawing and Fidelity Limits

- ## 9. Apply Quality Gate

+ Use a clean white control canvas with the same aspect ratio as P1. Draw one solid red (#FF0000) stroke, about 6–10 px at 1000 px width, without arrows, crossings, branches or extra coloured decorations. Place a green (#00FF00) start dot and blue (#0000FF) end dot beside the corresponding endpoints, close to the line but not painting over it. Use 3–6 clear bends. A loop needs a small gap with separate endpoint dots. These are screen-relative approximate directions, not GPS or a 3D trajectory: write altitude and look direction separately in Requirement. Up on the control canvas does not mean climb.

- Before returning a Design:

+ Local extraction must report verified direction before assigning P2 route movement. If missing, corrupt, branched, closed or unmarked, show ROUTE NEEDS REVIEW, retain the ground launch and 360-degree P1 orbit, then use the documented collision-safe FPV scene fallback. Do not invent P2 coordinates or endpoint. Preserve every overlapping verified path leg within each route-phase Shot, not just its midpoint. P2 pixels must never enter generation.

- - P1 is the only scene master; P2 is registered `analysis_only` and absent from visual loaders.

- - Shots cover the complete duration without gaps/overlap and use one main camera movement each.

- - There is exactly one continuous 360-degree orbit, one return to P1's exact original composition and one final one-second immutable-plate hold.

- - Landmark count, spacing and architecture remain stable; no collision or geometry flicker occurs.

- - Fireworks remain discrete sky particles behind/above buildings, with continuous smoke, reflections and exposure state.

- - Generated Pictures contain no orbit/yaw/trajectory clause in ordinary prompt text or keywords and carry both still contracts plus the dedicated negative prompt.

- - The final H3 prompt retains the true camera orbit but contains no route graphics, negative-prompt catalogue or visible control layer.

- - Multi-scene keyframes own disjoint ranges; future Pictures and previous audio never cross the wrong Segment boundary.

- - Native firework sound belongs to visible events and is not cut at Shot boundaries.

- - The terminal Picture is produced from P1's local pixels with effect-only compositing; it is never a text-to-image reinterpretation of the city.

+ Actual P1-conditioned img2img is best-effort, not pixel-exact multi-view reconstruction. It preserves the source framing as a supporting scene reference. Do not promise recovery of hidden building sides or exact path execution by a text-conditioned video model. Inspect P3 and every five-second scene reference before Preview. If P1 already contains a ring, request a clean P1; ordinary img2img may retain that artifact.