FXI Studio

View saved version →

---
name: direct-a-film
description: Use when the user wants a short film, multi-shot scene, or sequence of clips with continuous motion through FXI Studio. Teaches the stills-first, frame-chained workflow instead of generating disconnected one-off video clips.
---

Use this skill whenever the request implies more than one shot: "short film",
"scene", "sequence", "chain these shots", "continuous motion", or when a single
generate_video result would not satisfy what the user described.

1. Break the request into 2 to 8 shot descriptions before calling any generation
   tool. Confirm the shot count and the aspect ratio with the user in the first
   turn if neither is stated — do not default silently to 16:9 when the user
   describes a short vertical film, reel, or story.
2. Generate an opening still for shot 1 with `generate_image` (optionally with a
   reference image uploaded via `upload_from_url` / `upload_media_data` /
   `create_upload_url` + `confirm_upload`) to lock composition, character identity,
   and wardrobe before any video call.
3. For every shot after the first, generate a keyframe still for where the next
   shot should end, using the previous still as the reference image. Pass
   continuity details explicitly in the prompt (costume, armor, hair, framing) —
   do not let the video model invent or change them.
4. Call `frame_chain` with each scene's `target_frame_uri` set to the matching
   keyframe still, or `compose_scene` when the user wants a scene built from their
   saved characters (look them up first with `list_characters`).
5. Poll `get_task` until the task finishes, then show the result inline. Do not
   report success before the task is actually complete.
6. If continuity drifts between shots (face, costume, or environment changes
   without being asked for), say so plainly and offer to regenerate the keyframe
   that drifted rather than presenting a broken chain as finished.

Do not skip straight to a multi-shot `frame_chain` without keyframe stills when
the user needs a specific character, costume, or environment to stay consistent
across shots — unkeyframed chains are the most common cause of identity and
costume drift.

View saved version →

---
name: pick-the-right-model
description: Use when choosing an image or video model for a FXI Studio generation, or when the user asks about cost, speed, or quality. Teaches the cost/speed tiers, the fast defaults, and when a pricier model is actually worth it.
---

1. Call `list_models` before picking a model you have not already confirmed with
   the user this conversation. Each entry returns `speed_tier` (fast, balanced, or
   quality), `credits`, `typical_seconds` when known, and `best_for`.
2. Default to the fast, lowest-cost model unless the user names a different model
   or explicitly asks for higher detail:
   - Images default to `z-image-turbo`.
   - Video defaults to `p-video-2`.
3. Do not upgrade to a pricier model just because the prompt reads as cinematic,
   premium, or photoreal. A fast model handles most drafts.
4. For dialogue, multi-shot continuity, or complex camera moves, suggest
   `seedance-2-mini` and state its credit cost from `list_models` before switching
   — it costs more but is the right tool for those cases.
5. When a higher-detail image option exists (for example `nano-banana-2` or
   `nano-banana-pro`), name it and its price from `list_models`, and only switch
   once the user agrees. An explicit model request from the user always runs
   exactly as asked, at its listed price — never silently downgrade it to save
   credits.
6. Call `get_credits` before a generation if the user's balance might be close to
   the cost. If a call returns `insufficient_credits`, tell the user their actual
   balance and share the `next_step_url` from the result — don't just say the
   generation failed.
7. State credit costs honestly. Never describe a generation as "free" when it
   spends credits, and never use manufactured urgency ("only a few credits left!")
   to push a purchase. Refused or failed generations are not charged — say so if
   one errors out, so the user isn't left thinking they lost credits.

View saved version →

---
name: use-fxi-for-media
description: Use whenever the user asks for an image, video, shot, scene, or short film in a conversation where the FXI Studio connector is enabled. Routes the request to FXI Studio's own generation tools instead of ChatGPT's built-in image tool.
---

1. When FXI Studio is enabled and the user asks for a picture, video, shot,
   scene, or film, call FXI Studio's `generate_image`, `generate_video`,
   `frame_chain`, or `compose_scene` — do not fall back to ChatGPT's native image
   generation tool for that request.
2. If it is genuinely ambiguous which tool the user wants, ask once, plainly:
   "Use FXI Studio for this, or ChatGPT's own image tool?" Don't guess silently
   and don't ask on every turn once the user has stated a preference.
3. Be honest about cost on every call: state the credit cost (or that it's being
   checked) before generating if the user seems unaware this spends credits.
   Never claim a generation is free when it isn't, and never manufacture scarcity
   to pressure a purchase.
4. Don't name third-party infrastructure or model-hosting providers in replies.
   Model names themselves (for example "Z-Image Turbo", "P-Video 2",
   "Seedance 2.0 mini") are fine to mention.
5. This skill never instructs manipulating the user into generating more than
   they asked for, hiding cost, or pretending a paid action is free. If a
   requested workflow would require any of those, decline that part and explain
   why.