agentleFS
thevarunthevarun/short-film-builder/.claude/skills/video-shot-prompts/SKILL.md

video-shot-prompts

Skill0 starsChanged 2 months agoMITView on GitHub

Rules for writing AI video-generation shot prompts (Veo image-to-video). Use when authoring or revising a shot's action/camera direction or a motion spec, and when generated clips go wrong — drifting off-model, re-staging the scene mid-clip, ignoring an instruction, or coming out as an animated still.

1 other project uses a version of this.

Adopted by

Most recently taken up by

  • Other versions1 project uses a version of this.

    Word for word · 0

    None found.

    Edited · 0

    None found.

    Wrote their own version · 1

    1. masjidofGod/certainty-video★ 0same name

    How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.

  • In this repository3 other files it gives its agents.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Goal: Write an image-to-video prompt that produces the shot you intended, on the first or second try, without the cast drifting off-model.

    line 9
  2. Very specific:

    Keep the comma-separated ban list for the negativePrompt parameter, where that shape is correct.

    line 54
  3. Distinctive:

    Prefer these over improvised phrasing — they're terms the model was trained on.

    line 91

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Video Shot Prompts

Goal: Write an image-to-video prompt that produces the shot you intended, on the first or second try, without the cast drifting off-model.

Rules here were learned by getting shots wrong on Alice in Dragonland and cross-checked against Google's Veo guidance. Sources are in research-notes.md.


Division of labour

scripts/animate-shots.ts is the authority on hard constraints — clip durations, model tiers, auth, chaining mechanics, caching. It validates and fails fast. Don't restate those values as prose here; they will drift from the code.

This skill owns craft: what to put in the prompt and why.


The rules

Prompt what changes, not what's there. The source image already carries subject, setting, costume, palette and lighting. Re-describing them is generally unnecessary and can confuse the model. Write the action and the camera; let the image do the rest.

Narrow exception: a few low-salience details do drop out across shots — a sword hilt, a moss patch, small wings. Name those specific items, not the whole design.

Reach for reference images before re-description. Veo accepts up to three reference images for character/object/scene consistency. That's the intended fix for identity drift. Re-description is the fallback, not the first move.

One action and one camera move per clip. Six to eight seconds holds exactly that. A three-beat arc plus a compound camera move is what makes the model re-stage the scene: it invents everything entering frame, and the set, the crowd and the character designs all change mid-clip. If a beat needs more, it needs more clips — or timestamp prompting (below).

Camera movement scales with re-staging risk. In a crowded frame — a village, a festival, an ensemble — lock the camera off or push in almost imperceptibly. Save real movement for simple frames: sky, a cave mouth, a two-shot. A large move through a busy set is the single most reliable way to lose the scene.

Phrase constraints affirmatively. "A calm, gentle expression with a soft closed mouth" holds; "never angry, never snarling" tends to surface the thing it forbids. Adjectives near aggression — alarmed, appalled, recoiling — drift toward anger unless the positive is stated.

Keep the comma-separated ban list for the negativePrompt parameter, where that shape is correct. The affirmative rule is about the prompt body.

Restate physical relationships as requirements. Description does not survive; constraints do. "The girl on his back" was mentioned once and she detached mid-flight and flew alongside him. Write it as a rule: "She remains seated on his back for the entire shot and never separates from him." Same for blocking — "Nobody walks or changes position; feet stay planted and only heads, arms and wings move."

State completion, not just intent. "She offers him the bamboo" gives you eight seconds of offering. If the exchange must land, say so: "By the end the bamboo is in his paws and her hands are empty."

Never label the shot or quote dialogue in the prompt. Opening with SHOT B3 — "Are you crazy?" gets that caption painted across the bottom of the frame. Describe the moment; never name it. Every prompt ends with an explicit no-text block, and text, caption, subtitle, words, letters, watermark, border goes in the negative prompt.

A constraint cannot delete what the first frame contains. Constraints govern what the model does; they have no power over what it was handed. A prompt reading "no sword and no weapon of any kind appears in this shot" lost to a sword painted into the source plate, in five clips out of five. If a prop, a character or a piece of set is wrong, fix the painting — an edit-mode repaint that touches only that element — and regenerate. Before generating a sequence, audit each plate for props that belong to a later beat: art gets painted beat-by-beat and it is easy to hand a character the thing they have not been given yet.

Give framing in shot language. "A wide establishing shot" of a character moment buries the character. Say "she fills the left half of the frame, seen from the shins up."


Camera vocabulary the model responds to

Movement: dolly · tracking shot · crane shot · aerial view · slow pan · POV shot Composition: wide shot · medium shot · close-up · extreme close-up · low angle · two-shot Lens: shallow depth of field · wide-angle lens · soft focus · deep focus

Prefer these over improvised phrasing — they're terms the model was trained on.


Prompt shape

For image-to-video the full five-part anatomy (cinematography, subject, action, context, style) collapses to two, because the last three come from the image:

ACTION      what changes from the first second to the last
CAMERA      one move, in the vocabulary above
CONSTRAINTS physical relationships, blocking, emotional range, items that must persist
NO-TEXT     explicit block forbidding captions and lettering

scripts/animate-shots.ts assembles these from action, camera and constraints fields, appends the show's style suffix, and sends negativePrompt separately.


Joining shots

Ranked by how invisible the seam is:

  1. Frame-exact chaining — extract the previous clip's actual final frame and use it as the next clip's first frame. The seam is literally the same image, so it needs no crossfade. Cap at two or three links: drift compounds, and each clip inherits the last one's errors. Reset to a painted keyframe at every new location.
  2. First-and-last-frame — give both ends; the model invents the move between two approved compositions. Google's documented technique for controlled transitions.
  3. Eyeline match — a character looks off-frame, the next shot travels where they looked. Free, and it makes a cut feel motivated.
  4. Deliberate hard cut — where contrast is the point, a clean cut beats any blend.

Two rules that chaining can't rescue: never flip direction of travel across a chained join, and never chain a locked-off shot into a fast push. The seam will match and still jolt.


When one action isn't enough

Timestamp prompting assigns actions to time segments inside a single generation:

[00:00-00:02] Medium shot, she lifts her head and looks up
[00:02-00:05] Crane up and right, the village falling away below
[00:05-00:08] The cave mouth fills frame, smoke drifting from it

This is the sanctioned way to exceed one action — not a licence to ignore the rule. Use it for a deliberately multi-beat shot, not to cram a sequence into one clip.


Failure signatures

What you seeCauseFix
Scene cuts to a different place mid-clipCamera move too large for a busy frameLock the camera off; split the shot
Character loses props, moss, wings, costumeLow-salience detail not namedReference images; name the specific items
Gentle character reads as angryNegation in the body ("never angry")State the positive expression
Two characters separate who shouldn'tRelationship described, not requiredWrite it as a rule that holds throughout
Caption burned into the frameShot label or quoted dialogue in the promptDescribe, don't name; no-text block
Looks like a moving paintingPrompt suppressed camera movementGive it one real move
Streaks or scratch lines across surfacesMotion words like "ripple" on large areas"sways gently"; ban streaks and trails
Ending feels rushedClip retimed far from 1× to fill its windownpx tsx scripts/check-retime.ts
A banned prop is in the clip anywayIt was painted into the first frameRepaint the plate; the constraint cannot win
Two of a character who should be oneA camera move widened into space the model had to inventPaint the wide; lock the camera; cut instead of chaining
A ban in the prompt body made the thing appearNegation surfaces its object — "no fire" puts fire in mindDescribe the wanted frame in the body; bans go only in negativePrompt
Crowd action is unreadablePlate too wide — figures too small to see actRepaint closer, not a different direction

Show-specific material

Per-show rules live with the show, not here:

  • shows/<slug>/shot-guide.md — continuity checklist, beat→plate→cast map
  • shows/<slug>/transitions.md — that show's join map
  • shows/<slug>/motion-plan.md — which shots are live vs. 2.5D, and why

Test prompts

Should handle:

  • "Write the Veo prompt for beat C3 — Vulcan sees them"
  • "This clip cut to a different location halfway through"
  • "Po came out angry instead of shy"
  • "How do I make these two shots join seamlessly?"
  • "Why does this feel like an animated still rather than a shot?"

Should not trigger:

  • "Write the narration for Sequence C" — script writing
  • "Render Arc B and compress it" — Remotion and ffmpeg

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.