Global-AI-Group-Inc/content-generation-skills / skills/mira-generate
mira-generate
VerifiedGenerate images and video on Mira AI through its MCP server: connect, pick a model, attach the user's AI Blogger, location or product, write the prompt, pass a playbook skill, wait for the result, extend or upscale a clip. Use it whenever the user wants to "generate", "make a video", "make an image", "animate this photo", "create an ad", "сгенерируй", "сделай видео", "сделай картинку" and a Mira MCP server is connected or can be. Chain with mira-video-prompting and mira-image-prompting for what to write, and with the mira-* playbooks for the genre. NOT for prompting theory on its own, and not for voice, music, sound effects, dubbing or captions (mira-audio).
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 20.3 KB |
| references/ai-blogger.md | markdown | 14.1 KB |
| references/direct-presets.md | markdown | 3.6 KB |
| references/film-studio.md | markdown | 14.3 KB |
| references/guides.md | markdown | 9.3 KB |
| references/locks-and-references.md | markdown | 12.2 KB |
| references/marketing-studio.md | markdown | 5.6 KB |
| references/media-ops.md | markdown | 9.3 KB |
| references/miracle.md | markdown | 9.8 KB |
| references/playbooks.md | markdown | 1.8 KB |
| references/social-publishing.md | markdown | 4.9 KB |
references/locks-and-references.md
View on GitHub# References, locks and failure handling
## Brand assets
`list_brand_assets` returns three kinds of asset with ids:
- **AI Blogger**: the user's recurring character - a person (`kind: human`) or a designed
character or anthropomorphic animal (`kind: mascot`, which keeps its own render style in every
scene). Pass `bloggerId` (`avatarId` is the older name of the same parameter and still works).
The platform injects the canonical shots and an appearance passport (CHARACTER LOCK). Identity
never changes: face, hair, build, age, marks. Clothes, and the medium of a photoreal person, may
change; say so in the prompt when they should. A user with none can build one right here:
`create_blogger_draft` from traits or words, then `save_blogger` (ai-blogger.md); the AI Bloggers
section of the Mira app does the same. The character sheet (the 2×3 collage) is attached by the
platform itself; never pass it as a reference URL, a grid in the input turns into a split frame.
- **location** (`studioId`): a place. The platform injects its reference and a scene passport
(SCENE LOCK). Keep its signage and lettering in frame; they are the user's own, not third-party
brands.
- **product** (`productId`): the thing being sold. Its photos become references, its name and
description become facts (PRODUCT LOCK). The product stays the hero and keeps shape, colours and
label. Packaging may appear beside it, never over it.
Pass the id and describe the asset consistently. Do not paste the passport text into the prompt;
the platform already did.
## Reference URLs
`referenceImageUrls` takes up to eight public HTTPS URLs. One image on a video call turns it
into image-to-video with that frame as the start. Several images are merged into one start frame
on most models; `seedance-2-5` takes them natively and is the only family where an attached
AI Blogger or photo of a person reaches the model as an identity rather than a rebuilt frame.
Seedance screens every photo for real-looking faces before billing. When it refuses one, the
platform registers the refused photos in its private asset library for the length of that one
render, sends the same take again with them as assets, and deletes them as soon as the render
ends - the face stays the one in the photo, and nothing is needed from you: the call simply
takes a little longer. Only AI Blogger, uploaded and library photos go this way; a
stylised request (`style` anime, render3d or pixel, or a drawn medium in the prompt) is redrawn
into one composed start frame instead, because an asset would drag the clip back to photoreal.
Private and loopback addresses are refused.
An image attached to a chat is not a URL. If the client has file access, read the file and call
`upload_reference_image` with its bytes; if it does not, ask the user for a link.
## Keyframe anchors
`startImageUrl` and `endImageUrl` on `generate_video` are the clip's literal FIRST frame and the
frame it must land on. This is how shots chain without a visible seam: the end frame of one shot
becomes the start frame of the next.
They are a SEPARATE MODE at the provider, not two more references. With a pair attached you may
not pass `referenceImageUrls`, `referenceVideoUrls`, `referenceAudioUrls`, `imageRole`,
`bloggerId`, `studioId` or `productId` in the same call - ModelArk answers *"first/last frame
content cannot be mixed with reference media content"*, and the platform refuses the call before
a single credit is spent rather than letting you pay for a request the provider will reject. Make
two calls when you need both: one to build the frame, one to animate it.
A target end frame is accepted by `seedance-2-5`, `kling`, `minimax`, `omni` and the whole Veo
family; `list_models` reports `supports_last_frame` per model, and everything else refuses with
`end_frame_not_supported`. On most of them `endImageUrl` needs `startImageUrl` alongside it - a
clip cannot land on a frame without starting from one. `minimax` is the exception: it takes an
end frame ALONE, and the clip then starts wherever the model likes and simply has to land there.
`list_models` flags that as `supports_last_frame_only`.
Each family carries the anchors differently, which matters only when something goes wrong:
`seedance-2-5` and `wan3` send them as separate content roles, Veo as a `lastFrame` field, and
`omni` as plain words inside the prompt ("start exactly on photo 1 and land exactly on the last
photo") - Omni 1.2 deprecated its control tags. On `omni` a first-and-last-frame clip runs 3-10 s.
The end frame is a strong directional guide, not a pixel-perfect constraint: the clip moves
towards it and the final rendered frame may differ slightly. Build the second frame by EDITING
the first (`generate_image` on it) rather than writing a fresh prompt, or the model spends the
shot reconciling two different rooms.
## Video references
`referenceVideoUrls` on `generate_video` takes public HTTPS clips (mp4/mov, up to
200 MB each). The model borrows their motion, camera work, pacing or grade - never the people,
the place or the product in them. Only `seedance-2-5`, `seedance`, `minimax`, `wan3`, `wan3-prime`, `wan`,
`kling` (one clip, 3-10 s, Kling 3.0 Omni: `@video_1`, native audio off) accept clips; `omni`
does NOT - its video port is closed on our key;
every other model refuses with `video_not_supported`. How many and how long differs per model -
`seedance-2-5` takes ten of up to 30 s, `wan3` five of 15 s, `kling` one of 3-10 s - so read
`max_videos`, `video_ref_max_seconds`, `video_ref_min_seconds` and
`video_ref_max_seconds` from `list_models` instead of assuming three. Bind each clip in the prompt the way the model's file says: `@Video1`
on Seedance, `Video 1` on Wan, `reference video 1` on MiniMax, `<VIDEO_REF_0>` on Omni - an
unbound clip is ignored. A clip this account generated earlier (`list_generations`) is a valid
URL, which makes "do it again with this motion" a one-call job.
Seedance (ModelArk) screens reference clips the way it screens photos: a clip that shows a
real-looking person who is not one of the account's AI Bloggers is refused before billing
(`moderation_blocked`). Blockout playblasts, capsule or mannequin figures, stylised characters and
the account's own AI Bloggers pass. For a real stranger's motion take `minimax`, `wan3` (or
`wan3-prime`), `omni` or `kling-motion`. Give every reference ONE narrow job with an explicit
rejection, the way the providers' own guides put it: "@Image1 controls only the product's shape,
materials and label - do not copy its background, lighting or angle"; "@Video1 sets only the camera
path, the framing and the cuts - do not copy its figures, surfaces or colours". An image and a
video never do the same job.
`imageRole` on `generate_video` says what a SINGLE photo is for: `reference` - it shows the hero, the
place or the product and the model builds its own first frame (`seedance-2-5`, `seedance`, `veo*`,
`minimax`, `wan*`, `omni`, `kling`); `start_frame` - the photo is the literal opening frame the clip grows
out of; `auto` (default) - a reference on every model that has the mode, a start frame on `runway`,
`happyhorse`, `grok`. So when the user wants THIS photo animated (a product shot coming alive, a
live cover), pass `start_frame` explicitly. Two or
more photos are always references. Say which one you mean: a product shot you want animated is a start
frame; a photo of a place you want the clip to be set in is a reference.
`kling-motion` is the exception: it REQUIRES exactly one clip of a person performing the motion
(3-30 s, one person, single take, no cuts) and exactly one reference image of the character,
and the output is as long as the clip. The prompt only dresses the scene - wardrobe, setting,
mood - never the motion or the camera. When the character is the user's AI Blogger, take
`blogger_motion` instead: it builds the full-body frame for the clip's ratio (ai-blogger.md).
## Audio references
`referenceAudioUrls` on `generate_video` takes tracks (mp3/wav) the model borrows a voice, timbre
or rhythm from. `seedance-2-5` takes up to ten and `wan3`/`wan3-prime` up to five; `list_models`
reports `max_audios` per model, and everything else refuses. The provider needs at least
one photo or clip attached alongside - audio on its own is refused
(`audio_requires_visual_reference`). Bind each in the prompt as `@Audio1`, the same way clips are
bound: an unbound track is ignored.
## References beside a continuation
`extend_video` takes `referenceImageUrls`, `referenceVideoUrls` and `referenceAudioUrls` too, on
the models that accept references beside a continuation. `seedance-2-5` takes photos, clips and
audio; `omni` takes PHOTOS beside its continuation (a photo attached to an extend turn really is
used - a car photo put the car in the continuation) but no clips; Veo continues from the footage
alone and refuses references rather than dropping them silently.
The numbering shifts by one: the source clip itself is `@Video1`, so your clips start at
`@Video2`, and one fewer fits than on a fresh render. Their seconds count towards the provider's
total video budget TOGETHER with the source, so a long source leaves little room - trim the
references rather than sending more of them.
## Continuing backward
`extend_video` with `direction="backward"` (`seedance-2-5` only) writes a prequel: the new segment
is stitched IN FRONT of the clip. The source is still `@Video1`, but now it is where the new shot
must end, not where it starts. Write the moments before the clip opens and land on its first frame:
the same subjects, poses and framing at the instant the clip begins. Do not describe what the clip
itself shows. The price and the steps are the same as forward; `list_models` reports
`extend_directions` per model, and a forward-only model refuses rather than extending the wrong way.
## Drafts and finals
`generate_video` with `draft=true` on `seedance-2-5` renders a 480p take for a fraction of the price.
Iterate on drafts, then `finalize_draft` the one the user picks: the provider re-renders THAT take in
1080p with the same seed, prompt, references and sound. It is a new render, not an upscale, so fine
texture, tiny text and crowds can shift slightly. A draft can be finalized once, within 7 days, and
cannot be extended or upscaled; a 1080p final extends in 1080p at the 1080p price. Drafts do not
combine with keyframe anchors (`startImageUrl`/`endImageUrl`).
When the user already knows the shot, skip the drafts: `quality="hd"` renders `seedance-2-5`
straight to 1080p at the same 1080p price a final costs (about 100 credits per 5 s). Drafting
first pays off only when there are several takes to choose from. `list_models` reports
`supports_hd`.
Every model with more than one resolution takes the same `quality` parameter; the tier a model
offers is in its row of [selection.md](../../mira-video-prompting/references/selection.md).
`sd` is the tier below the standard one (720p on `kling`, `kling-turbo`, `kling-motion`, `minimax`,
`wan3`, `wan3-prime`, `happyhorse`; 360p on `omni`) for cheap tries, `ld` is 480p on `wan3` and
`wan3-prime`, `hd` is 1080p on `kling-4`, `seedance-2-5` and `omni`, `uhd` is 4K on `kling` and
`omni`. Each tier has its own price - quote it with `estimate_cost` - and a tier the model lacks
is refused before any charge. `runway` renders at one resolution.
## Seeds
Most models ignore the seed. Reuse it together with a lightly edited prompt for a close
variation; a model change discards it.
## When a call fails
- 401 on `tools/call`: the sign-in was not completed. Ask the user to finish it in the browser.
- `out_of_credits`: say so plainly and stop. Do not retry with a cheaper model unasked.
- `moderation_blocked`: the provider refused the input. A refused face on `seedance-2-5` has already
been retried once as a temporary asset and, failing that, as a composed start frame; a
refusal after both means the photo will not pass -
describe the person instead of attaching it. Do not resubmit the identical call in a loop:
the face screen is not deterministic, but a second identical call costs the user the wait
again. A refusal on reference CLIPS is final: see above.
- `still_running` from `wait_for_generation`: call it again. Video routinely needs three or four
calls.
- An unknown model id does not error; it falls back to the registry default. Check with
`list_models` when unsure.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |