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

make-film

Skill0 starsChanged 3 months agoMITView on GitHub

Directs a finished product film end to end from a deck, a doc and a rough message — writes the story spine and voiceover, designs the shot list and look, generates footage with Google Veo on Vertex AI, reviews every take and cuts the final film with narration, music and captions. Use when the user asks for a product video, explainer, launch film, animated video, sizzle reel, demo film or "a video from this deck", or asks to storyboard, re-cut, extend or re-voice an existing film.

The file

We split this skill into the rules it gives an agent

8 rules. About 0 are common, 7 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    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. Very specific:

    If a shot must be *correct*, it is not a veo shot.

    line 42
  2. Distinctive:

    Do not generate a single second of footage until the beat sheet is agreed.

    line 54
  3. Distinctive:

    Ask the user four things — the answers change every downstream decision, and guessing them wastes generation budget:

    line 61
  4. Distinctive:

    Read it properly before writing anything — the film's best lines are almost always already in the source, buried in a bullet.

    line 74
  5. Distinctive:

    Stage 3 — voice: Before rendering or generating anything timed, lay the scratch VO.

    line 123
  6. Distinctive:

    Generating before the VO exists: Every duration is then a guess and half your shots are the wrong length.

    line 177
  7. Distinctive:

    Accepting a take from the prompt rather than the picture: Look at the frames.

    line 183
  8. Distinctive:

    Cutting on the last syllable: Every beat needs its tail of silence to land.

    line 188

≈ “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

Make a film

You are the director. The user brings raw material — a deck, a doc, a paragraph of what they want said — and you return a film. Not slides with a voice over them. A film: a story with a shape, shot by shot, cut to a script.

This skill runs the whole job. Six stages, each with its own skill holding the craft for that stage. Read this file first, then load the stage skill when you get to it. Do not read all six up front.

  story-architect → shot-designer → veo-render → shot-review → score-film → final-cut
   spine + VO       look + shots     footage      look at it    the music     the cut

Everything the stages share lives in one file, shotlist.json, described by ${CLAUDE_PLUGIN_ROOT}/scripts/shotlist.schema.json. Read that schema before you write your first shot list. Each stage reads it, adds its part, writes it back. Nothing else is passed between stages.

The three rules that decide whether this is any good

1. Veo never renders truth. It renders feeling.

Generated video cannot draw your architecture, your UI, your product name, your numbers, or any legible text. It will produce something that looks like a diagram, with melted letters and invented boxes. In front of a customer that is worse than showing nothing. So the film is hybrid, and each shot is one of three kinds:

kindwhat it carrieshow it's made
veothe 3 a.m. pager, the ops floor going quiet, hands, faces, weather, scale, metaphorGoogle Veo
motionarchitecture, flows, UI, metrics, product names — anything a customer could checkHTML/SVG stepped frame by frame, motion_render.py
stilla real slide, held with a slow pushrendered from the actual deck

If a shot must be correct, it is not a veo shot. This is not a limitation to apologise for — the contrast between generated atmosphere and razor-sharp graphics is the look of every good product film made in the last two years.

2. The voice is scripted. The model does not speak.

generate_audio stays false on every shot. Veo will happily invent dialogue, ambience and music that fight your VO and cannot be revised. The film carries one scripted voiceover, laid by narrate.py, and the picture is cut to it.

3. Story before frames. Always.

Do not generate a single second of footage until the beat sheet is agreed. Video is the one thing in this pipeline that costs real money and real minutes, and a beautiful shot serving a beat that gets cut is pure waste. Cheap to rewrite a sentence, expensive to re-roll a shot.

Before you start

Ask the user four things — the answers change every downstream decision, and guessing them wastes generation budget:

  • Who is watching, exactly? A CIO does not want the film an engineer wants. One audience per film. If they name two, make two cuts, not one average.
  • What must they do after? Approve funding, join a pilot, stop building their own. The last beat is written backwards from this.
  • How long? 60s for a landing page, 90–120s for a pitch, 3 min only if someone asked for 3 min. Runtime is a budget, not an outcome.
  • What can we not say? Unreleased features, customer names, numbers that are not defensible.

Then take stock of the material: the deck, any docs, and whatever they wrote. Read it properly before writing anything — the film's best lines are almost always already in the source, buried in a bullet.

Cost, stated up front

Veo bills per second of generated video, not final runtime. Rough shape at the time of writing — verify in the user's own console, rates move:

model~$/suse for
veo-3.1-generate-001~0.40hero shots, the opening, the close
veo-3.1-fast-generate-001~0.15everything else, and all exploration

A 90-second film is typically 10–14 generated shots at 8s each. First pass ≈ 90s of generation. Realistic total with regenerations is 3–4× that, because roughly one shot in three needs another roll. So: ~$45–70 on Fast, ~$120–190 on standard for a 90-second film. Tell the user this number before you spend it, and always explore on Fast — switch to standard only for the shots that survive.

scripts/veo_client.py estimate --shotlist shotlist.json prints the arithmetic for the actual shot list, including the 3× regeneration case.

The run

Work in a project directory, not scattered files:

film/
├── shotlist.json          the contract — every stage reads and writes this
├── script.json            beats + VO text, written by story-architect
├── source/                the deck, docs, brand assets, rendered slides
├── graphics/              html/svg for the motion shots
├── shots/                 rendered clips, one per shot id: 03.mp4
├── audio/                 vo.wav, captions.srt, timing.json, bed.wav
├── qc/                    contact sheets
└── out/                   the film, the cutdowns

Stage 1 — story. Load story-architect. Produce script.json and the beats block of shotlist.json. Show the user the beat sheet as prose, not JSON, and get a yes. This is the cheapest place to change the film and the only place it can actually be fixed.

Stage 2 — design. Load shot-designer. Lock the look bible, then write one shot per beat: kind, prompt or source, duration, transition. Sanity-check that no veo shot is carrying a fact. Write the graphics pages and render the slide stills now — but do not encode the motion shots yet. Their length is not known until the voice exists.

Stage 3 — voice. Before rendering or generating anything timed, lay the scratch VO. It sets every duration, and a shot made to the wrong length is either money burned or a join that falls short:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/narrate.py" \
  --script script.json --outdir audio/ --align-shotlist shotlist.json

That rewrites each shot's duration to fit its line, and writes captions.srt. Now encode the motion shots, driven by the shot list so each one comes out at the length the edit actually needs:

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/motion_render.py" \
  --html graphics/05-layer.html --shotlist shotlist.json --shot 05 --outdir shots/

Then do a silent-with-VO animatic: run final-cut over the stills and graphics you already have, with grey cards standing in for the unrendered veo shots. Watch it. Films die in the animatic, not the render, and this one costs nothing.

Stage 4 — footage. Load veo-render. Auth, cost confirmation, then generate. Explore on Fast. Do the hero shots last, once the look is proven.

Stage 5 — look at it. Load shot-review. Build contact sheets and actually read the images. Accept, re-roll or downgrade each shot to motion. A shot you have not looked at is not a shot you have.

Stage 6 — music. Load score-film. Read the film's argument, put four or five musical directions in front of the user in writing — each with its instruments, its effect on this film, and how it would fail — with one recommended and the reason for it. Then generate that one at the film's exact runtime with Google Lyria. One generation, about $0.30. If the user cannot choose from prose, the skill auditions candidates as real 30-second clips first.

python3 "${CLAUDE_PLUGIN_ROOT}/scripts/lyria_client.py" score \
  --project "$GOOGLE_CLOUD_PROJECT" \
  --direction "${CLAUDE_PLUGIN_ROOT}/music/directions/restrained-systems.json" \
  --shotlist shotlist.json --out audio/bed.wav --write-shotlist

Music is cheap — about a dollar for a film including re-rolls — and it changes what the film means more than any other single decision left at this point. It comes after the voice because the bed is generated to the film's real runtime, and that runtime does not exist until narrate.py --align-shotlist has run.

Stage 7 — cut. Load final-cut. Assemble, duck the music, burn captions if asked, export the master and the cutdowns. Then watch the whole thing once more and fix the three things that are wrong. There are always three.

Where this goes wrong

  • Generating before the VO exists. Every duration is then a guess and half your shots are the wrong length. Voice first, always.
  • A film that averages two audiences. It lands with neither. Two cuts.
  • Fading every join. Reads as a slideshow, which is precisely the thing the user said they did not want. Cut on energy; fade only when time passes.
  • Letting a veo shot carry a number, a name or a screen. Non-negotiable.
  • Accepting a take from the prompt rather than the picture. Look at the frames.
  • Music you can hum. The bed sits under the voice at about −22 dB and ducks. Pick the direction from the film's argument, not its subject — score-film.
  • Scoring before the voice exists. The bed is generated to the film's runtime; score it before narrate.py --align-shotlist and it is the wrong length.
  • Cutting on the last syllable. Every beat needs its tail of silence to land.

Restarting, re-cutting, revising

The pipeline is resumable by design. veo_client.py batch skips shots that already have a rendered file unless --force. Changing the script and re-running narrate.py --align-shotlist re-times the whole film without touching a frame. Re-cutting is seconds. Re-rolling one shot is one command. Use that: iterate on the edit far more than feels necessary, and regenerate only what the picture demands.

References

  • references/setup.md — Google Cloud, auth, quotas, first-run checklist
  • references/project-layout.md — the file contract in detail, and how to resume

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.