Skip to content

Global-AI-Group-Inc/content-generation-skills / skills/mira-generate

mira-generate

Verified

Generate 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).

references/miracle.md

View on GitHub
# Miracle

Miracle changes anything in a clip while every move stays: put the user, a friend or an AI Blogger
in place of the people in a video, swap an outfit, a product, the place or a sign, redraw the whole
clip in a style, add people who were not there, and let the blogger say a line at the end. The
camera, the timing and the sound stay as in the original. On the website it is the Miracle tab of
AI Bloggers (next to Builder and Speak). The name stays "Miracle" in every language.

It works on clips of 3 to 30 s, on one of two engines; `miracle_analyze` says which ones a clip can
use and what each costs:

- **One pass** (Seedance 2.5 video edit, the default): the whole 4-30 s clip in a single render,
  720p, up to 30 photos, about 5 minutes. Offered for any clip of 4-30 s when the platform has it
  switched on; a clip or a photo the engine refuses for a real face is retried automatically, and
  only if that fails does the run go in parts (the price difference comes back). Styles and added
  people exist only here.
- **In parts** (Kling 3.0 Omni): a clip longer than 10 s is cut into parts of up to 10 s (at scene
  cuts where it has them), each part is rendered on its own and the parts are joined back under the
  original sound. 1080p, up to 4 photos per part, real people's photos allowed, about 5-8 minutes.

`engine` on `miracle` is `auto` (default: one pass when the clip and the cast allow it), `seedance`
or `kling`. Whether one pass takes photos of real people is in the analysis (`real_person_photos` of
that engine); when it does not, a real person's photo sends the run in parts.

## Tools

| Tool | Credits | Use it for |
|---|---|---|
| `list_miracle_presets` | free | The motion library by `shelf`: `preset` (platform clips with fictional people, each already analysed), `community`, `hero`. |
| `miracle_analyze` | free | Reads a clip: the parts, every person with a position and a description, the objects that can be swapped, the styles, the price. Lives 24 hours. |
| `miracle` | per second of the clip, by engine | The run itself, on an analysis. One pass: 10 per second (720p); in parts: 6 per second (1080p). A line adds the speech and 1 per second of lip sync. |
| `estimate_cost` | free | kind `miracle`, `sourceSeconds` = the clip length, `model` = `seedance` or `kling`, `chars` = the line. |

## Two modes

- `transfer` (motion transfer): the clip's motion, camera and timing are kept; the cast, the place
  and the look are rebuilt from the references. Typical: "me and my friend doing this dance",
  "this walk, but in Shibuya at night".
- `swap`: only what is named changes - a person, an outfit, a product, the place or a sign; the
  rest of the frame stays as filmed. Typical: "replace the can in her hand with our bottle",
  "change the sign to say FRESH BREAD".

## Flow

1. Source. A preset (`presetId` from `list_miracle_presets`), a clip the account made
   (`sourceGenerationId`) or a file sent through `create_upload` (kind `video`, then
   `libraryItemId`). 3 to 30 s.
2. `miracle_analyze` with that source. Show the user what was found, in their language and in
   plain words: "1 · on the left - young man in a dark grey cardigan; 2 · on the right - ...",
   the objects, the length and the price of each engine it offers (one pass or in parts). A person
   seen only in some parts is marked with those parts.
3. Ask who becomes whom and what changes. For each person: keep, a saved AI Blogger
   (`bloggerId` from `list_brand_assets`) or a photo (`photoUrl`: `upload_reference_image` or a
   Mira image). People not listed stay as they are.
4. Say the price and wait for a clear yes. Then `miracle`, and `wait_for_generation`.

```
miracle_analyze {"sourceGenerationId": "<dance clip>"}
miracle {"analysisId": "<from miracle_analyze>", "mode": "transfer",
         "cast": [{"personId": "p1", "bloggerId": "<the user's AI Blogger>"},
                  {"personId": "p2", "photoUrl": "<https photo of the friend>"}],
         "locationUrl": "<https photo of a night street>", "consent": true}
miracle {"analysisId": "<...>", "mode": "swap",
         "swaps": [{"target": "product", "objectId": "o2", "imageUrls": ["<https photo of the bottle>"]},
                   {"target": "text", "objectId": "o3", "text": "FRESH BREAD"}]}
```

## Rules

- In parts, at most 4 photos go into one part of the clip: the people visible in that part, the
  new place and the swapped objects; a person who is not in a part does not count there. In one
  pass the limit is 30 photos for the whole clip. Over the limit is refused
  (`miracle_too_many_images`); asking for one pass where the clip or the cast does not allow it is
  refused too (`miracle_engine_unavailable`).
- `swaps[].target`: `outfit`, `product` or `location` with 1 to 4 `imageUrls`; `text` with the
  new lettering in `text` (up to 80 characters). `objectId` comes from the analysis; a swap
  without it applies to every part.
- `locationUrl` is for `transfer`; in `swap` the place is a swap with `target: "location"`.
- `prompt` (optional, up to 400 characters) adds a direction: light, weather, a style. On its own
  it is enough for a `transfer` run that only restyles the clip.
- `keepSound` (default true) keeps the original sound under the result.
- A photo of a real person needs `consent: true`: ask the user to confirm the person is them or
  agreed to it. Never a public figure, never a face the user has no consent for. A run with
  nothing to change is refused (`miracle_nothing_to_change`).
- An analysis expires after 24 hours (`miracle_analysis_expired`): analyse again. Analyses are
  free but limited to 30 per hour.
- Each run is one generation; failed runs return the credits in full. The result is a regular
  video generation: `extend_video`, `upscale_video`, the sound tools and `draft_mira_guide` take
  it like any other clip.

## What Miracle can and cannot change

- The ground the motion touches stays: stairs, a ramp, a bench someone sits on, the shape of a
  skatepark come back in the new place as the same shapes. Pick a new place that can host the
  same path, or tell the user what will remain.
- A swapped product takes the pose of its photo and the motion of the original object; it does
  not gain a rotation the original never made. Shoot or pick the product photo at the angle the
  original object shows most.
- A small object can change shape (a sneaker into a bottle); a large moving one only changes its
  skin. A car swapped for a bus or a trolley comes back as the same car repainted: swap vehicles
  within their class.
- When the place changes, the platform already names the cars, signs, street lamps and other
  lights of the old place that the analysis found, so they do not survive. Anything else that must
  go - a parked scooter, a bench - goes into `prompt` as "where the scooter stood there is a palm
  tree".
- If a product was not swapped in a run that also changed the people or the place, run a second
  `swap` on the result (the result is a regular generation) instead of re-rolling the whole run.

## Words for `prompt` that backfire

The model reads words, not intent. "Cart" becomes a car - write "supermarket trolley". "Security
guard", "staff" or "checkout" summon police. "Frozen" turns people into statues - write what they
do instead. Any uniform or armband written without a plain colour and "no insignia of any kind"
can come back with a symbol on it. A line in quotes summons subtitles; the platform adds the ban,
so keep quotes for words that must really appear on a surface.

## Style, added people, a line

- `style` redraws the whole clip, every move, cut and position kept: `anime`, `cartoon_3d`, `clay`,
  `comic`, `watercolor`, `oil_painting`, `sketch`, `pixel_art`, `noir`, `retro_film` (the analysis
  lists them). `styleImageUrl` takes the medium, palette and line work of an image instead (or on
  top), never its people or place. It goes with `cast` and `addPeople` in the same pass.
- `addPeople`: up to 2 people who were not in the clip, each `bloggerId` or `photoUrl` (a real
  person's photo needs `consent`), a `position` (`left` ... `right`, `background`) and an `action`
  ("dances along", "claps"). They get the same light and a believable size.
- Both work only in one pass: with them a clip that cannot go one pass is refused
  (`miracle_seedance_required`), and `engine: "kling"` is not allowed.
- `lineText` (with optional `lineVoiceId`, `lineLanguage`): after the pass the blogger says the line in
  their pinned voice, lip-synced; the clip's sound is dropped. It needs exactly one person in frame
  after the edit (background extras do not count), that person a blogger or given `lineVoiceId`, and
  a photoreal look (no style, or `noir` / `retro_film`), and it must fit the clip at about 2.5 words a
  second. If the lip sync fails, the video comes without the line and that part of the price comes back.

## Own background

`performerBloggerId` (with `performerFraming` `full` or `half`) is the other way to use a clip: the
blogger repeats its motion in their own frame and place (Kling 3.0 Motion Control) instead of being
put into the clip's scene. Mode `transfer` only, without `cast`, `swaps`, a place, a style or added
people; `lineText` works. `blogger_motion` with `engine: "kling"` does the same without an analysis.

## Miracle or another tool

- One person into one 3-10 s clip, no analysis needed: `recast_video` or `blogger_motion` (mode
  `recast`) is enough.
- A clip with cuts, several people or many references (a cast of AI Bloggers, a product and a new
  place): Miracle in one pass keeps every cut in a single render.
- A blogger repeating a dance in their own frame: `performerBloggerId` here, or `blogger_motion`
  with `engine: "kling"`.
- Several people, a clip up to 30 s, or a person plus a place or product at once: Miracle.
- One object in a clip up to 10 s: `swap_object` works too; Miracle does the same on longer clips
  and several objects.
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: 472kUpdated Claude CodeCodexVerifiedInstall
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: 323kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 323kUpdated Claude CodeCodexVerifiedInstall
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 availableStars: 323kUpdated Claude CodeCodexVerifiedInstall