Skip to content

moonlight-lupin/agent-skills / creative/clips-studio

clips-studio

Verified

Use when the user wants to "make a video/clip", "generate a video from text", "animate this image/photo", "create a marketing reel / social video / teaser", "do a 3D / parallax move", "pan/zoom/orbit a shot", or mentions fal.ai, Kling, Veo or Seedance for video. Staged fal.ai workflow — brainstorm, draft cheaply, produce the final at quality. Three modes: text-to-video, animate (still to motion), camera-move (push-in/pan/orbit). Not for still images (image-studio), slide decks, or charts. Honesty discipline: never depict real identifiable subjects via text-to-video.

FileKindSize
README.mdmarkdown1.9 KB
SKILL.mdskill md23.5 KB
examples/example-run.mdmarkdown8.1 KB
references/fal-video-models.mdmarkdown11.9 KB
references/motion-honesty-guide.mdmarkdown11.4 KB
references/video-prompting-techniques.mdmarkdown3.4 KB
scripts/falvid.pyscript34.5 KB
tests/__init__.pyscript0 B
tests/test_falvid.pyscript31.3 KB

references/motion-honesty-guide.md

View on GitHub
# Motion honesty guide — animating real subjects without misrepresenting them

A moving clip is more persuasive — and therefore more capable of misleading — than a still. This
guide is the discipline that keeps an AI clip honest. Read it before Stage 1; it's the video
counterpart to `image-studio`'s faithful-edit discipline. The discipline **scales with how real the
subject is** — a real place, a real product, real people. **Property / real estate is a common worked
example below**, but the same rules apply to any real subject.

**Text-to-video invents the entire scene**, so never use `generate` to depict a real, identifiable
place, product or person — `animate` a real photo instead. The rules below assume a real subject; for
a purely fictional/abstract brand-mood clip the honesty bar is lower, but never imply it shows
something real.

## Contents

- The one rule
- Animating people (`animate`)
- Hands & fine object interaction — the believability ceiling
- Screens, TVs & projected content
- Camera moves invent geometry (`camera`) — the big one
- Extending a frame (outpaint to change aspect)
- Never
- When this skill is the wrong tool
- Label & review

## The one rule

> **Show motion and atmosphere; never invent space, features, or facts about the subject.**

An AI video model *hallucinates* everything it can't see in the start frame. That is fine for
**how light moves** or **how a space feels** — and dangerous when it starts inventing **what the
place is**: rooms that don't exist, a view that isn't there, a finish the subject doesn't have, a
crowd that implies an occupancy it doesn't have.

## Animating people (`animate`)

People bring a space to life, but they also imply things. Keep it **illustrative lifestyle**, not a
documentary claim:

- **Do** — natural, relaxed activity at believable density: a couple of people chatting, someone
  reading, a person walking through. Genuine interaction reads better than posed smiling.
- **Don't** — imply a specific **demographic, occupancy level or activity** the subject doesn't
  actually have (a packed, buzzing room for a half-occupied space; a particular clientele).
- **Don't** — animate **identifiable real people** (likeness + egress). Generated, non-identifiable
  figures only; never upload a private person's photo.
- **Faces & hands deform in motion** — prefer fewer, larger, mid-distance figures; if a face warps,
  reduce the crowd and re-draft rather than chasing it.

## Hands & fine object interaction — the believability ceiling

The single hardest motion for **every** current image-to-video model (Kling, Veo, Seedance alike) is
**hands articulating around small objects** — pouring or holding a kettle, gripping a cup, typing,
handling a phone, working a door handle. Models tend to **exaggerate the gesture** and **morph the
hand or the object** mid-action. A better / production model *reduces* this but does **not** eliminate
it — so **don't pay for the quality render to fix hands; fix the brief on the cheap model first.**
(Observed repeatedly in testing: a sofa pair told to "chat" came out with unnatural, over-large
gestures; restraining them to small micro-motion fixed it, while a background figure asked to use a
kitchenette kept "pouring" something unconvincingly. On a busier frame the ceiling was harder still: a
*focal* seated person with a snack bowl kept raising a hand toward their face across **three**
re-drafts — even with the bowl parked aside, hands placed in the lap, and explicit negatives. Text has
a hard limit against a determined hand-to-face prior on a focal figure.)

**The rule: design the hard motion *out* — don't rely on the model to nail it.**

- **Park object-handling figures as static**, explicitly: *"stands facing the counter, hands at their
  sides, glancing down; does not touch, lift, hold or pour anything — no cup, no kettle."* Naming the
  non-action works better than hoping the model skips it.
- **Keep foreground people small and restrained** — a slight nod, a small smile, a gentle shift of
  weight; **hands resting, no large gestures.** Name the action *small*. Restrained motion also reads
  more premium, and is the more honest frame.
- **Negatives are a soft steer, not a guarantee** — include
  `exaggerated gestures, morphing or distorted hands and fingers, warping objects, fast jerky motion`,
  but it's the prompt *parking the action* that does the real work, not the negative.
- **Distance forgives.** A far/background figure's imperfection is far less obvious than a focal one —
  so put any unavoidable fiddly motion in the background and keep the **focal** subject's hands
  simple (resting, still, no object).

This is the practical face of the skill's "**a better prompt before a pricier model**" rule (SKILL.md,
Stage 2): re-draft cheaply with the motion designed down, *then* spend on the final.

### The fallback — when people won't behave, move the camera instead

If a **focal** figure's hand/object motion still won't behave after a cheap re-draft or two, **stop
animating the people and move the camera instead.** A **camera move** (people held, only the camera
drifts) sidesteps the hand-to-face and object problems entirely, and its motion *masks* residual
jitter in tricky elements (a fidgety on-screen TV, a far figure). Observed: a slow camera push-in held
a struggling figure — and everyone — still where three `animate` re-drafts couldn't.

**This is a fallback, not a default.** The default for a "bring it to life with people" brief stays
`animate`; reach for the camera move only when people-animation hits this ceiling. (The `camera` mode
remains a *first-class* mode in its own right for camera/3D briefs you actually want — it's just not
the first answer to a struggling people-clip.) And if a clip genuinely *needs* believable object
handling, it's the wrong tool altogether — use a real shoot.

## Screens, TVs & projected content

A screen in frame (a TV, a cinema/projector screen, a monitor) is its own trap: **any "alive" cue
makes the model fabricate motion on the screen** — it will cut to a different scene, then zoom, rather
than hold the image. That's both unnatural (it reads as a glitch) and a mild fabrication (invented
on-screen content).

- **Keep screen motion out of the brief.** Don't ask the screen to "play"; if anything, say *"the
  image on the screen stays the same, a faint steady glow; it does not change scene, cut or play new
  footage."*
- **Even "stays the same" is not a hard lock** — the model may still drift or zoom the screen
  (observed: cut scenes on one draft, zoomed on the next, despite the instruction). Don't keep
  fighting it with more prompt.
- **A camera move masks it.** Under a slow camera push-in the screen moves *with* the frame, so
  residual screen drift is far less obvious than when the camera is locked — another reason the camera
  fallback suits screen-heavy rooms.

## Camera moves invent geometry (`camera`) — the big one

A camera move asks the model to render **what the camera would see as it moves** — i.e. parts of the
scene that **were never in the photo**. The more the camera travels, the more the model invents.
Ranked from safest to riskiest:

| Move | What it invents | Verdict |
|---|---|---|
| **`--static`** (locked camera, subject moves) | almost nothing — only the moving subject | **Safest.** Prefer it when you just need life/motion. |
| **Slow push-in / pull-out** (small dolly) | a little edge detail; mild parallax | **Good.** Gentle and largely faithful. |
| **Slow pan / tilt** (small angle) | whatever rotates into frame at the edges | **OK if small.** Keep the angle modest. |
| **Shallow orbit** (a few degrees) | the near side of objects as they turn | **Caution.** Stay shallow; a wide orbit fabricates whole new faces of the space. |
| **Deep fly-through / big orbit / "walk the room"** | entire walls, rooms, the far side of a façade | **Avoid.** This is no longer the real place — it's an invented one. |

Practical rules:

- **Prefer gentle, slow, short.** A 5s slow push-in sells "premium and calm" and invents almost
  nothing. A sweeping orbit invents a building.
- **Speed and travel are prompt-driven** — there's no numeric speed control on these models, so you
  have to *spell out* the pace (a bare `--move push-in` tends to come out faster/larger than
  expected). **Match the pace to the scene** — a calm interior usually wants a slow, subtle drift; a
  façade or hero reveal can carry a more deliberate move. The only fixed point is the honesty
  trade-off: the faster and farther the camera travels, the more edge geometry it invents — so pick
  the *slowest move that still does the job*, not a one-size pace.
- **Use `--static` whenever it does the job** — it's the most honest motion there is.
- **Watch the draft for hallucination**: doorways that lead nowhere real, a window gaining a view,
  furniture multiplying, a corridor inventing an end. If the move reveals space the photo never
  showed, pull the move back.
- **Façades / exteriors**: a slow push-in or a small lateral drift is fine; orbiting a building
  invents elevations you don't have photos of.

## Extending a frame (outpaint to change aspect)

To fit a different aspect — a vertical photo into a 16:9 hero, say — the cheapest route is to
**outpaint** the still wider in `image-studio`, then animate. But outpainting **invents the
surroundings that weren't photographed** (the buildings next door, more street and sky). For a real
subject that's a misrepresentation risk, so:

- **Ground it in reality.** Ask the user for **reference images of the actual surroundings** —
  street-view, the neighbouring buildings, a wider shot — and use them so the extension *reconstructs*
  the real context instead of inventing it. Grounded beats invented, every time.
- **Confirm the extended still before animating.** It's a cheap image step before the video spend —
  show it, take corrections (*"that's a building, not empty space"*, *"trim the sky"*), re-edit, get
  approval, **then** animate. Don't animate an unconfirmed invented frame, and remember the invented
  context is then *in motion* — doubly illustrative.
- **Keep it illustrative.** Never present an outpainted clip as the real street or setting; label it.
- **Prefer a properly-oriented source photo** over outpainting at all (see SKILL.md, Stage 1).

## Never

- Never **add or improve** a feature, finish, fixture, view or amenity the subject doesn't have.
- Never **remove** a real, permanent feature to flatter the space.
- Never present the clip as a **measured tour, a real walkthrough, a floor-plan-accurate
  representation, or a record of the actual layout**. It is an **illustrative clip**.
- Never animate **confidential or sensitive content**, or an **identifiable real person** without
  consent.

## When this skill is the wrong tool

For anything where **literal accuracy matters** — a property listing, an offer document, a spec claim,
a measured tour — an **AI extrapolation is not appropriate**. Use:

- a **real video shoot** of the actual subject, or
- a **proper 3D tour / Matterport-style capture** built from real measurements.

This skill is a fast first-pass **marketing** draft for screens, posts and reels — clearly an
illustration, reviewed by a person, never a document of record.

## Label & review

- Treat every output as a **draft for review** by a person before any use.
- Where the clip will be published, recommend a visible **"illustrative" / "for marketing purposes"**
  note in the post or the asset itself (the skill produces the clip; the human decides distribution).
canvas-designanthropics/skillsCreate beautiful visual art in .png and .pdf documents using design philosophy. You should use this skill when the user asks to create a poster, piece of art, design, or other static piece. Create original visual designs, never copying existing artists' work to avoid copyright violations.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
docxanthropics/skillsUse this skill whenever the user wants to create, read, edit, or manipulate Word documents (.docx files) or Word templates (.dotx files). Triggers include: any mention of 'Word doc', 'word document', '.docx', '.dotx', or requests to produce professional documents with formatting like tables of contents, headings, page numbers, or letterheads. Also use when extracting or reorganizing content from .docx or .dotx files, inserting or replacing images in documents, performing find-and-replace in Word files, working with tracked changes or comments, or converting content into a polished Word document. If the user asks for a 'report', 'memo', 'letter', 'template', or similar deliverable as a Word or .docx file, use this skill. Do NOT use for PDFs, spreadsheets, Google Docs, or general coding tasks unrelated to document generation.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
pdfanthropics/skillsUse this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text/tables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting/decrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
pptxanthropics/skillsUse this skill any time a .pptx or .potx file is involved in any way — as input, output, or both. This includes: creating slide decks, pitch decks, or presentations; reading, parsing, or extracting text from any .pptx or .potx file (even if the extracted content will be used elsewhere, like in an email or summary); editing, modifying, or updating existing presentations; combining or splitting slide files; working with templates (.potx), layouts, speaker notes, or comments. Trigger whenever the user mentions "deck," "slides," "presentation," or references a .pptx or .potx filename, regardless of what they plan to do with the content afterward. If a .pptx or .potx file needs to be opened, created, or touched, use this skill.Stars: 198kUpdated Claude CodeCodexVerifiedInstall