agentleFS
FSBuilder-mayuresh93FSBuilder-mayuresh93/product-film-director/skills/shot-designer/SKILL.md

shot-designer

Skill0 starsChanged 3 months agoMITView on GitHub

Turns an approved beat sheet into a locked look and a shot list — deciding which beats become generated Veo footage, which become deterministic motion graphics, and which are stills, then writing the Veo prompts, durations and transitions. Use when storyboarding a film, writing or fixing Veo prompts, choosing shots for a script, or when generated shots look like six different films.

The file

We split this skill into the rules it gives an agent

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

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

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

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    If the audience must read it, it is a motion shot.

    line 99
  2. Very specific:

    Do not set durations by taste — narrate.py --align-shotlist will overwrite them to fit each VO line.

    line 122
  3. Distinctive:

    Run the check yourself, out loud, against the finished shot list:

    line 140

≈ “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

Shot designer

Two jobs, in this order: lock the look, then assign every beat a shot. Doing them the other way round is how you end up with twelve beautiful clips that do not belong in the same film.

1. The look bible

Written once into shotlist.json.look, then obeyed by everything. It is short on purpose — a paragraph you could hand a DP.

"look": {
  "prompt_suffix": "shot on 35mm anamorphic, shallow depth of field, cool teal-slate palette with a single warm practical, soft volumetric haze, slow deliberate camera, muted desaturated grade, cinematic corporate documentary",
  "negative_prompt": "text, letters, words, logos, watermark, subtitles, captions, user interface, charts, diagrams, distorted hands, extra fingers, morphing objects, fast cuts, cartoon, 3d render, stock footage look",
  "palette": ["#0B1F2A", "#1C4E5C", "#E8B04B", "#F4F6F7"],
  "type": {"family": "Inter", "weight": "600", "case": "sentence"},
  "grade": "eq=saturation=0.94:contrast=1.05",
  "reference_images": ["source/style-01.png", "source/style-02.png"]
}

Four levers do almost all the continuity work, in descending order of power:

  1. reference_images — up to three, passed to every Veo shot as style references. This is the single strongest tool you have. Generate one shot you love, or take a frame from the deck, and reference it everywhere after.
  2. prompt_suffix — identical camera, lens, grade and lighting language on every prompt. Never vary it shot to shot. Vary the subject; hold the world.
  3. grade — one ffmpeg filter across all shots at edit time, which quietly pulls generated and rendered footage into the same skin.
  4. negative_prompt — global, and it always includes text, logos, watermarks, captions and UI, because Veo adds them uninvited and they are unfixable.

Derive the palette and type from the user's actual deck, not from taste. Pull hex values from the source file if you have it.

2. Assign every beat a kind

Work down the beat sheet and ask one question per beat: could a viewer check this? If yes, it is not veo.

veo — people, places, weather, hands, scale, time passing, abstract motion, anything felt rather than verified. The pager at 3 a.m. The empty ops floor. A city at dusk seen from a datacentre window. Light moving across a server aisle.

motion — architecture, data flow, UI, metrics, names, timelines, anything with a number or a word in it. Built as HTML/SVG and stepped frame by frame with motion_render.py, so it is exactly right and repeatable. Rebuild the geometry from the deck; do not screenshot a slide and call it a graphic unless you mean to.

still — a real slide, held 4–6s with push: 1.05. Honest, instant, and the right answer for a proof beat where the slide is already good. Nothing on screen is ever truly static; always give a still its slow push.

A healthy 90-second film is roughly half veo, a third motion, the rest still. All-veo looks impressive and says nothing. All-motion is a corporate deck with easing.

3. Write the Veo prompts

One shot, one continuous camera move. There are no cuts inside a prompt — if your prompt contains "then", you have written two shots.

Structure every prompt the same way:

[subject and setting] · [action, one verb] · [camera: move, framing, lens] ·
[light and time of day] · [mood]

Worked example:

A lone network engineer at a standing desk in a dark open-plan office, lit only by three monitors and one desk lamp; she rubs her eyes and leans back as the screens keep refreshing. Slow dolly-in from behind her shoulder, medium wide, 35mm. Pre-dawn blue through the window, hard cyan monitor spill. Exhausted, quiet, no drama.

Then look.prompt_suffix is appended automatically. Do not repeat the suffix's content in the shot prompt.

Rules that actually change the output:

  • Say the camera move explicitly. "Slow dolly-in", "locked-off", "gentle handheld drift", "crane down". Left unsaid, Veo picks something busy.
  • One action. Two actions in eight seconds reads as a glitch.
  • Describe light, not colour grading. "Pre-dawn blue through venetian blinds" beats "cool colour grade" — the suffix handles grade.
  • Never name a brand, product or real person. It will be rendered as garbled text or an uncanny lookalike.
  • Never ask for text or UI. Even "a screen showing dashboards" tends to produce hallucinated interface. Say "screen glow", "unreadable data reflected in glass". If the audience must read it, it is a motion shot.
  • Avoid crowds and hands in close-up unless the shot needs them. Both are where generation still visibly breaks.
  • Generate 8s even when you need 5. Trimming is free; the last second of a generated clip is often where it falls apart, and the extra headroom lets you cut on the good frames.

Continuity between shots

  • image — set the first frame. Feed it a rendered graphic and the generated motion inherits your exact geometry and colour. This is the best trick in the toolkit for bridging motion into veo.
  • last_frame — set both ends and let Veo interpolate. Use it to hand one shot into the next invisibly, or to land on a frame the next graphic starts from.
  • extend_from — continue an approved clip rather than re-rolling a longer one. Cheaper and preserves the take you already liked.
  • seed — pin it once you have a take you like, then vary the prompt slightly around it. Record the seed of everything approved in the shot's render block.
  • samples: 2 on hero shots only. Each extra sample bills like a full generation.

4. Duration and rhythm

Do not set durations by taste — narrate.py --align-shotlist will overwrite them to fit each VO line. What you do set is duration_seconds (what Veo generates: 4, 6 or 8) and the transitions.

Transitions carry meaning, so spend them:

  • cut — the default for energy, contrast, and any beat that answers the one before it. Most joins in a good 90-second film are cuts.
  • fade, 0.4–0.6s — time passing, or a change of world (generated → graphic).
  • A film that fades every join is a slideshow. That is the exact thing being avoided here.

Give the film one deliberate breath: a single beat around two-thirds through with no VO, just picture and music. It is the moment the audience realises they are watching something made, not assembled.

5. Before you spend anything

Run the check yourself, out loud, against the finished shot list:

  • Does any veo shot carry a fact, name, number or screen? Fix it.
  • Does every beat have a shot, and every shot a beat?
  • Is generate_audio false everywhere? It must be.
  • Is look.prompt_suffix on the plan for every veo shot?
  • Render the motion and still shots now — they cost nothing. Cut the animatic with grey cards for the veo gaps and watch it before generating anything.

Then run the estimate and show the user the number:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/veo_client.py" estimate --shotlist shotlist.json

References

  • references/veo-prompting.md — the prompt grammar in depth, twelve worked prompts for corporate/technical subjects, and the failure catalogue
  • references/motion-graphics.md — building motion shots: the HTML harness, window.__seek(t), easing, and matching deck geometry exactly

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.