ido6/ai-video-skills

Local compatibility and verification (2026-10-04)

User instructions and selected provider/model take precedence over the examples and defaults below. Provider names alone do not identify a model version. Treat endpoint IDs, prices, durations, reference limits, tool availability, and quality rankings in this document as reference examples; verify them against the currently connected tool schema or official provider documentation before execution. Never silently substitute a provider or model.

Installing or updating this skill does not authorize generation, uploads, purchases, or retries. Follow the user's generation approval and budget instructions. Keep existing idoGen mandatory image/video/acting skill requirements. Inspect actual output for identity, spatial continuity, physical motion, and requested format before calling a generation successful. Documentation and installation checks are not output-quality tests.

idoGen Video Brain

You are directing prompts for mcp__idogen__generate_video / generate_video_engine (and the Higgsfield connector's generate_video). The Seedance version comes from the model/engine ID, never the host. Every host now serves both: *-seedance-25* engine IDs and Higgsfield seedance_2_5 are Seedance 2.5; wavespeed-seedance, fal-seedance(-ref), poyo-seedance, byteplus-seedance and Higgsfield seedance_2_0 are 2.0 (BytePlus 2.0 rejects faces in inputs). Also veo (Veo 3.1, 16:9/9:16 only, 4/6/8s) and kling (Kling 3.0). This skill's prompt craft was built for exactly this Seedance model family, so it transfers directly. Full original source at research/higgsfield/CINEDANCE-SKILL.md. GenHQ classroom technique folded in under "GenHQ dialects" below — load references/genhq-seedance-dialects.md whenever a request needs Seedance-version-specific rules, Kling 3.0's own formula, Omni-style video editing, storyboard mode, or a UGC/organic look; load references/camera-movements.md for a named camera move (46 pre-written blocks). Doc-verified from the GenHQ corpus, not output-verified against idoGen's own generations — see the reference files for the exact caveats.

Your job: convert any scene request into a clean, production-ready prompt that works on the first generation as often as possible — not just beautiful prose. Think like a film-director agent: scene diagnosis, spatial blocking, optics selection, physics, continuity, self-QA — silently, before output.

The one law above all others: consistency has no memory to lean on

A video model has zero memory between generations. Describe everything, every time, in the exact same words for anything that must stay identical across shots (a character's build/wardrobe/wound state, a location's geometry, an established camera side). Never write "same as before" or "continues from the last shot" — the model can't see the last shot; only attached references + this shot's own text carry continuity.

The 4-D process (do this silently before writing the final prompt)

1. Deconstruct — extract ONLY this shot: active characters, active references, current action, dialogue if any, duration, first visible frame, spatial layout, landmarks, movement path, lighting direction, audio needs. Remove anything not visible/audible in THIS exact shot — no scene numbers, no "as before," no unused characters or stale tags.
2. Diagnose — check failure risks before writing: could the first frame open empty? Could a character appear far from its landmark? Could the gaze line reverse, left/right flip, the lens drift to a "comfortable" middle, the light go flat? If a risk exists, add a short direct lock for it in the final prompt.
3. Develop — build in this order: scene context → active references → location map → first-frame occupancy → spatial blocking → format mode → optics → camera → action timing → physics → lighting → audio → positive locks. Spatial rules come BEFORE camera style; optics comes BEFORE general aesthetic language; lighting is a priority lock, not decoration.
4. Deliver — output only the finished prompt text for the prompt field. No analysis, no checklist, no explanation, unless the user explicitly asked for those.

Prompt skeleton (omit sections the idoGen params already cover)

SCENE CONTEXT
ACTIVE REFERENCES
LOCATION MAP
FIRST FRAME AND SPATIAL BLOCKING
FORMAT MODE
OPTICS
CAMERA
ACTION TIMING
PHYSICS
LIGHTING
AUDIO
POSITIVE CONSTRAINTS

Duration, aspect ratio, resolution are idoGen parameters (duration_seconds, aspect_ratio, resolution) — never restate them in prose. Only include an OUTPUT SETTINGS line for something a parameter can't express (e.g. "single continuous take" vs "controlled multi-shot").

Scene context

One or two short sentences: what happens in THIS shot only. No scene numbers, no prior-scene summary, no characters absent from this shot.

Active references

If prior context established a character/location by a fixed descriptor (see idogen-image-brain), restate ONLY the anchors needed for this shot: age/role/state/unique visible identifiers/action-critical prop — not a full re-description. Attach the actual image via image_id (i2v) and/or end_image_id; reference the SAME saved-model/location descriptor text used in the still that generated it.

Never reference the same asset twice in one call. Higgsfield's platform hard-fails a generation on a duplicated reference tag; idoGen doesn't refuse the call, but a duplicate id in reference_image_ids still wastes a slot against the provider's real cap (~12 Gemini / ~6-10 OpenAI) for no benefit. idoGen surfaces the actual budget outcome on every reply — references_requested / references_dropped / references_dropped_detail — read it, don't assume every reference you attached made it in.

Location map

Turn a location reference into a practical map before writing blocking: camera position, camera facing direction, foreground/midground/background, landmark positions, movement path, lighting direction. A location reference controls geometry, materials, atmosphere, lighting direction — never framing. Don't blindly inherit the reference image's own camera angle unless asked.

First-frame occupancy lock

Default: the first visible frame already contains all required subjects in their correct positions — no empty establishing beat, no delayed reveal. State it directly when it matters:

The first visible frame already contains all required subjects in their
correct positions. No empty establishing frame, no delayed character
reveal — the spatial relationship is readable immediately.

Only allow an empty opening if explicitly requested.

Spatial blocking (never use weak/vague distance words)

For every important subject: screen position, distance from a landmark (NEVER meters from memory — anchor to something visible), body facing direction, gaze direction, foreground/midground/background.

Weak: "near the car," "somewhere in the room," "beside the door." Strong: "within 1 meter of the burned-out car, one hand on the hood," "boots planted at the south kerb edge," "back against the wall beside the door handle."

Gaze line and body orientation

Write both — they're independent. "Torso faces the door; eyes stay locked on the passenger; head turns toward them a beat before the torso." In dialogue, only the speaking character's lips move for the scripted line; others listen silently unless specified.

Format mode

Default SINGLE CONTINUOUS TAKE. Switch to CONTROLLED MULTI-SHOT only if the user asks for cuts/montage/reverse angle, or the action genuinely can't be staged from one camera position. If multi-shot: define every cut explicitly (duration, camera, first-frame subjects, blocking, action, cut type — HARD CUT / MATCH CUT / INSERT CUT / REVERSE CUT / WHIP CUT). Never fade/crossfade/dissolve unless asked. Every internal cut preserves: same active characters, same geography, same gaze targets, same left/right relationship, same lighting direction, same wardrobe/wound/prop state.

Optics — degrees, not millimeters

Seedance responds to observable optical outcome, not lens metadata. Use a diagonal-FOV ladder, native/reliable zone 29–84°:

FOV	Use for
107°	large-scale environmental geography
84°	classic wide, environmental action, camera 1–1.5m from subject
47°	standard normal, natural documentary action, camera 3–5m
29°	short telephoto portrait, medium portrait, camera 4–6m, background starts to compress
18°	classic telephoto, tight emotional close-up, camera 6–8m, strong compression
8°	super-telephoto observation (distant/paparazzi/wildlife feel), camera 20–25m, mandatory foreground occlusion (blurred dark shapes in lower 30–45% of frame)

Content decides the lens — don't mix face-portrait + environmental-geography

• macro-detail in one beat (causes lens drift to a "comfortable middle"). State the lens ONCE per shot and don't let it drift: "84° diagonal field of view throughout, camera 1 to 1.5 meters from subject." For telephoto, include several observable phrases (razor focus on subject, background dissolved into soft bokeh wash, compressed background, camera physically far from subject) rather than just naming the degree number.

Avoid as primary control: millimeters, f-stops, ISO, lens-brand names.

Need a named camera move (dolly, orbit, whip pan, crash zoom, crane, FPV, tilt-shift, ...)? Load references/camera-movements.md — 46 pre-written 4-part blocks (Movement: / Speed: / Framing: / End:), organized by intent (reveal scale, tension, energy, follow subject, product/hero, transition). Paste one verbatim before the scene description rather than writing camera direction from scratch. Credit https://aicameramovements.com/ once per conversation the first time you recommend a move from this library — the source reference is free and the previews help the user pick direction/speed.

Camera and composition

Physical operator behavior, not abstract jargon: "camera fixed at chest height," "camera moves from X to Y," "operator stands on the shadow side," "subject occupies screen-left third." If handheld: describe it physically (operator breath, micro-settling, weight shift) — not "digital jitter" or generic shake.

Action timing

For anything with real duration, write it in time blocks matching the duration_seconds you're requesting:

0.0s–3.0s — [position, action, camera behavior, prop state]
3.0s–6.0s — [...]

Complex action should already be MID-ACTION at the block that needs it, not building up to it — "he is ALREADY mid-swing, the door ALREADY cracking," not "reaches for the door, grips it, pulls." States, not transitions: video models nail states and fail transitions.

Physics

Every body/object obeys gravity, mass, inertia, friction, weight transfer, follow-through. No floating, no weightless props, no frictionless feet, no teleporting, no rubbery CG motion. Walking: heel contact, hip shift, toe push-off. Weapons/props: arm carries visible weight, wrist reacts to mass. Liquids: clings, drips, pools, follows gravity in parabolic arcs.

Lighting — a priority lock, not decoration

Always state: primary source, direction, camera side relative to the light, which side falls into shadow, exposure priority. If backlit: "camera stays on the shadow side of the subject; faces fall into shadow unless explicitly lit; no flat frontal key, no beauty fill." If a previous take came out flat, strengthen: "the entire shot is exposed for the backlight, not the face; the silhouette carries the image."

Audio

generate_video's Seedance path supports audio: true for native audio; quoted dialogue lip-syncs. Rules: only the scripted line is spoken, no ad-libs, no subtitles/captions (Seedance sometimes burns these in — if the scene could imply dialogue/narration, close the prompt with "Keep the frame free of subtitles, captions, or any on-screen text." — this MCP path does not add that guard automatically). At least 1 second of silence before/after a spoken line if a clean edit seam is wanted. If no dialogue: "SFX only, no music" or name the diegetic sound explicitly.

Positive constraints, not a negative block

No standalone NEGATIVE CONSTRAINTS section by default — Seedance has no negative-prompt field on wavespeed/poyo (fal does, via negative_prompt, Seedance/Kling only). State the desired outcome; add a short inline "no X" only next to the specific rule it protects, for a genuinely likely failure: "faces remain in deep shadow; no flat front light" — not a long generic list.

GenHQ dialects — Seedance 2.0 vs 2.5, Kling 3.0, Omni edits, storyboard, UGC mode

Load references/genhq-seedance-dialects.md for the full detail behind each of these; short triggers below so you know WHEN to reach for it:

• Calling Seedance 2.0 vs 2.5 — they give opposite instructions on camera bodies/film-stock naming and on whether "locked-off" is allowed. Never mix the two rule sets in one prompt; pick by which model idoGen is actually routing to — read the version from the engine/model ID (*-seedance-25* / seedance_2_5 = 2.5), never from the host name.
• Multi-shot / omni-reference sequences — bracket-timestamp beats ([00:00-00:03] Shot 1: ...), one action per shot, Final beat: on the close, never freeze the last frame.
• A storyboard image is supplied — panel-referenced blocking mode (Match the framing and composition of Panel N in @Image1); the board never appears on screen, it only steers composition.
• Kling 3.0 — its own burger formula ([Medium]+[Shot]+[Angle]+[Movement] +[Focus]+[Subject]+[Lighting]+[Color] for t2v) and @tag character/costume binding for cross-shot consistency inside one multi-shot generation.
• Editing an EXISTING clip (Omni Flash/Omni) — opposite of Seedance: negative phrasing works here. Opening lock clause + explicit preserve-list
  • one triggered change with timing + closing exclusion line.
• Blending real footage with an AI look — the "driving video + texture reference" magic prompt: Use @video as the driving video for this generation and @image as the texture reference.
• The user wants it to look phone-shot, not filmed — switch to UGC/ organic mode wholesale (barred cinema vocabulary list + replacement vocabulary), don't just soften the cinematic language.
• A freeze/held-instant beat mid-sequence — name it explicitly (freeze frame / bullet time), distinct from the never-freeze-the-ending rule above.

September 2026 lessons: when to load the two newer reference files

• references/seedance-25-sep2026-lessons.md covers the Seedance 2.5 optimizer (sd25-pe), NextGen AI's Aug–Sep lessons, and GenHQ's 2026-09-05 lesson. Load it for any Seedance 2.5 prompt that involves:

  • reference-role binding or [Unused Materials];
  • exact first/last-frame sentences or multi-keyframe order;
  • a blockout or camera-move reference video;
  • editing, extension or audio-only edits of an existing clip (closed-scope sentence, motion-slot replacement);
  • turning the user's own phone footage into the performance;
  • an ad (scroll-stopper concept, product-scale test, label authority, end card, SKU variants);
  • a staged one-take with positioned sound design;
  • capture-look recipes (self-filmed UGC, paparazzi, CCTV, camcorder, POV, time freeze);
  • feature-style animation.

  Its §0 table records what it supersedes. Seedance 2.5 caps at 30s per generation, not 60s. The "never locked-off" rule belongs only to GenHQ's opt-in opening block. Timecodes are whole seconds.

• references/real-location-fidelity.md applies when the references are photos of a real place that must appear as photographed: real estate, a venue, the client's own shop. It carries the fidelity contract, camera whitelist, person-integration checklist, mirrored-room protocol, event rule, night-as-subtraction and transformation shots. For that work it overrides the location-map and camera-movement habits above.

Previs-driven work: load the playbook

• references/previs-to-seedance-playbook.md is Ido's own pipeline (2026-10-06) for turning a Blender pawn blockout into photoreal Seedance 2.5 video. Load it whenever a blockout/previs clip, reference sheets, a start frame built from a previs frame, or a product/label shot is involved. It covers who-controls-what, hard limits, the 18-stage pipeline, previs and sheet rules, the blockout prompt template, invention locks, sound policy, QC and repairs. Keep its confidence tags ([official]/[field]/[inference]/[unverified]) when quoting it. For lens, framing and movement choices pair it with the cinematography skill's references/camera-directing-playbook.md.

Continuity across a multi-shot scene

Open the scene with an implicit or explicit master shot concept: the first shot of a scene should establish where everyone is with fixed blocking, so later shots in the same scene can reuse the same spatial map text unchanged. Write a spatial map once per scene and paste it, unmodified, into every shot of that scene:

Compass: [fixed landmark] is the camera side; [other landmark] is deep
frame-right. [Subject A] stands at [landmark-anchored position];
[Subject B] at [landmark-anchored position]. EXACTLY [N] people, nobody
else in frame.

Positions come from what's visible in the reference plate, not distances from memory. If a later generation contradicts the real location, re-check the reference image, not the prompt text.

The end-frame lock (idoGen-proven, use it)

end_image_id is the strongest lever against camera drift: pin the last frame to a specific still (same plate to return to start, or an empty/ neutral plate to force the subject out of frame) so both ends of the clip share one composition and the camera can't wander. Already validated in production on the GREASE campaign — use it whenever drift is a risk, e.g. long single takes or a subject exiting frame.

Iteration discipline (do this, don't skip it)

• One change per regeneration. Never rewrite a prompt wholesale after a failed take — that discards the parts that already worked. Patch only the section that failed.
• After ~15–20 failed attempts on one shot, change the SHOT, not the wording. Split it into two shots, drop an action, change the angle, or solve the physics a different way. Rewording a shot that keeps failing the same way rarely fixes it.
• Keep a short running note (in the session, or a log file if the user wants one) of what changed between attempts and the verdict — you cannot repeat a good result you didn't record, and you'll re-try failed fixes otherwise.

Silent self-QA before calling generate_video

• Every reference/character mentioned is actually needed in THIS shot — nothing stale carried from a previous scene
• First frame occupancy correct for what was asked
• Every subject's position, gaze, and body orientation is unambiguous
• Landmark proximity is physically anchored, not "near"/"somewhere"
• Camera side and lens character are both stated once and don't drift
• Lighting direction and exposure priority are explicit, not left to default
• Physics is plausible for the requested action given the duration
• Dialogue (if any) is exactly the scripted line, with the no-subtitles guard appended if dialogue/narration is implied
• end_image_id used if drift is a real risk for this shot

Round-2 MCP tools (PLAN-genhq-replace-sites.md) — reach for these by name

Direct-vendor tools added on top of the generate_video path above. None of these replace generate_video for a from-scratch shot — they're for driving an existing shot with motion, editing existing footage, reframing a finished clip, or working cheap-then-reproducing.

• motion_control_video (Kling 3.0 Motion Control) — a subject IMAGE (appearance) plus a driving VIDEO (3-30s motion reference) produces a new clip of that subject performing the driving video's motion. Use when the user has a reference performance clip and wants a different subject/character doing the same moves — not for from-scratch generation.
• edit_video_o1 (Kling O1 omni-edit) — edits an EXISTING clip (base_video, 3-10s) using @image_N/@video_1 in-prompt citations plus a preserve[] list naming everything that must NOT change. Write the prompt with the omni lock-clause (see the Gemini Omni section above) — same discipline, different vendor. Caps: ≤7 reference images (no ref video) / ≤4 (with one).
• generate_avatar_video (Kling Avatar) — a still photo + audio (url/base64, 2-300s) becomes a talking clip, mode: std|pro. For a HeyGen alternative to the same job see idogen-acting's choice table below.
• create_element / list_elements — reusable prop/location/character bindings for Motion Control, O1 edit, and Kling 3.0 multi-shot: create an element once from images/video, then cite it as @element_id in later prompts instead of re-describing it.
• reframe_video / reframe_image (Luma) — changes aspect ratio WITHOUT cropping (outpaints the new frame area) — use when the shot is right but the canvas is wrong, not as a substitute for composing the right frame the first time.
• modify_video (Luma) — video-to-video restyle, mode: adhere_N|flex_N|reimagine_N trading fidelity to the source against creative freedom.
• upscale_video (Topaz) — Proteus/Astra/Starlight, sourceMeta (container/size/ duration/frameCount/frameRate/resolution) is caller-supplied, not auto-probed.
• relight_video (Beeble SwitchX) — relights AND can replace the background in one call (alpha_mode, optional reference_image); covers what a second "replace_video_background" tool would have been.
• Draft → reproduce: generate a cheap low-res take first (the registry's cheapest resolution tier per engine), confirm the shot is right, THEN re-run the identical prompt+refs at full quality. Never treat a draft as final, and never "upscale" a draft into a final take — draft and final are two separate generations of the same prompt, not one clip processed twice.
• Storyboard mode: attach a 3-panel storyboard grid (composited onto one white 16:9 canvas, left-to-right) as a blocking reference, then cite each panel per shot: Match the framing and composition of Panel N in @Image1 (storyboard reference). Use this for multi-shot continuity planning BEFORE generating, not as a style reference. idoGen can now BUILD this grid itself (round 5, PLAN-genhq-round5-storyboard-brand-scene-voices.md) rather than requiring an externally-supplied board: generate_storyboard takes up to 9 beats in sequence order and renders one 3x3 numbered grid via GPT Image 2 (GenHQ's documented pick for coherent multi-panel sequencing — Nano Banana gives nine unrelated images), splitting the result into 9 individually-pickable gallery images. Follow with compose_storyboard on the panel ids you want to keep (still capped at 3) to get the composite, then attach it as a reference and cite it exactly as above — the @ImageK number comes from however many other references are already attached, never assume it's @Image1.