Productclaude-opus-5-5

A River Did This: the Grand Canyon in 57 seconds

A 57-second vertical film of the Grand Canyon being cut, made entirely in code. It opens on "A river did this.", starts on a flat plateau six million years ago, and the canyon deepens as a year counter runs down. Every fact is on screen as words, and it ends on a one-line card. Terrain, water, trees, sky, light, the score and the wind and water sounds were all generated by code; no photo, texture or sample was used.

Remix

How it was made

Reconstructed from the publisher's account.

Models

  • claude-opus-5-5 · anthropic · execution · Claude Code

Agent

Claude Code

Tools and APIs

  • Skill

    Load the Procedural Film Maker skill.

  • Agent

    Fresh reviewers scored each round's stills.

  • numpy

    Build and erode the terrain, make the sound.

  • scipy

    Shape the terrain and the sound.

  • moderngl

    Run the GLSL shaders on the GPU.

  • GLSL

    Ray-march the water, trees and sky for each frame.

  • ffmpeg

    Encode the film.

Ordered steps

  1. Fill in the skill's brief

    BRIEF.md in an empty folder that holds only the skill. These are the facts and the end line the film had to use.

    # Brief for the Procedural Film Maker
    
    ## What place should it build?
    
    The Grand Canyon, and the big story of it told in under a minute: a flat plateau, then a river cutting down through it over about five to six million years until it is a mile deep. Show the layered red and tan rock walls, the Colorado River at the bottom, the plateau rim. The film should feel like time-lapse of millions of years: the canyon deepening as we watch.
    
    Facts to put on screen (check them against a good source): about a mile (1.8 km) deep, about 277 miles (446 km) long, carved mostly by the Colorado River in about 5 to 6 million years, while the rocks in the walls are far older, up to nearly 2 billion years at the bottom.
    
    ## Reference photos or a mood (optional)
    
    No photos. Warm low sunlight, deep shadows in the canyon, not grey.
    
    ## The film: length, aspect, pacing, music
    
    45 to 60 seconds, vertical 9:16 for X and phones. The claim in big words on screen in the very first second (for example "A river did this."), and every fact also shown as words on screen, because most people watch with the sound off. Calm wind and water sounds and an epic score made in code, nothing bought. It must end on this one line on screen, exactly: "Email, logins and 2FA codes for each of your AI agents. actas.dev"
    
  2. Start the skill

    The one message that started the run. The agent worked, rendered and checked everything on the same laptop.

    Use the procedural-film-maker skill on BRIEF.md.
  3. Restart after a machine restart

    The machine running the session restarted mid-render. Nothing in the project changed.

    Your last command was killed (exit 137) when the machine running this session restarted. Nothing else changed. Carry on.
  4. Restart after a second machine restart

    The preview it had started finished while the session was down.

    The preview render finished: it wrote out/preview_540_round4.mp4 (log in out/cache/preview.log). Carry on.
  5. Answer the agent's question

    After four rounds with scores far below its bar, the agent asked: render now, keep iterating, or both. The answer chose render now, and fixed one number.

    Thanks, I watched the preview and it works for me. Go with option 1: render the final 1080x1920 film now and say in your notes that it's below the photoreal bar. One fix first: the card 'Today it is about a mile deep (1.8 km)' mixes two facts. A mile is about 1.6 km, and 1.8 km is the deepest point. Make it say 'about a mile deep (1.6 km)'. Render on this machine as you have been.

Made with the Procedural Film Maker skill

See the skill

US$19 for this skill.

Buy the Procedural Film Maker skill

One payment, then you download it. It runs on your own machine, in Claude Code or Codex, with your own subscription.

By buying, you agree to our Terms and Refund Policy. Your purchase is sold through Link, a Stripe company, which takes your payment.

About

A 57-second vertical film of the Grand Canyon being cut, made entirely in code. It opens on "A river did this.", starts on a flat plateau six million years ago, and the canyon deepens as a year counter runs down. Every fact is on screen as words, and it ends on a one-line card. Terrain, water, trees, sky, light, the score and the wind and water sounds were all generated by code; no photo, texture or sample was used.

How it was made

The Procedural Film Maker skill, run in Claude Code with Claude Opus 5.5, got the brief shown here and nothing else. It wrote a shot list, built the canyon as an eroded height map with ray-marched water and trees and a physical sky, and scored stills of every shot over four rounds with fresh reviewers. It checked the facts against the National Park Service and rendered the film on the same laptop in 58 minutes. Nobody edited the film by hand.

Recorded: made by the Procedural Film Maker skill in a clean folder holding only the skill and a brief, 12 November 2026.

What to know

  • This is below the skill's own bar. Reviewers scored the weakest part of each shot around 41 out of 100 against "indistinguishable from a photo". The rock walls read as smooth clay rather than fractured cliffs. After round 4 gained nothing, the agent stopped and asked whether to render; the person giving the brief said render, labelled below the bar.
  • The brief itself said "a mile (1.8 km)". 1.8 km is the deepest point; a mile is 1.6 km. The agent noted the difference, and the person giving the brief had the card changed to "(1.6 km)" before the final render.
  • The skill has since been told to stop after two rounds without gain and render without asking, and to check every number and unit on screen.

Exact prompt

Hidden characters, if present, appear as Unicode escapes
# Brief for the Procedural Film Maker

## What place should it build?

The Grand Canyon, and the big story of it told in under a minute: a flat plateau, then a river cutting down through it over about five to six million years until it is a mile deep. Show the layered red and tan rock walls, the Colorado River at the bottom, the plateau rim. The film should feel like time-lapse of millions of years: the canyon deepening as we watch.

Facts to put on screen (check them against a good source): about a mile (1.8 km) deep, about 277 miles (446 km) long, carved mostly by the Colorado River in about 5 to 6 million years, while the rocks in the walls are far older, up to nearly 2 billion years at the bottom.

## Reference photos or a mood (optional)

No photos. Warm low sunlight, deep shadows in the canyon, not grey.

## The film: length, aspect, pacing, music

45 to 60 seconds, vertical 9:16 for X and phones. The claim in big words on screen in the very first second (for example "A river did this."), and every fact also shown as words on screen, because most people watch with the sound off. Calm wind and water sounds and an epic score made in code, nothing bought. It must end on this one line on screen, exactly: "Email, logins and 2FA codes for each of your AI agents. actas.dev"

What you need

  • Claude Code with the Procedural Film Maker skill
  • Python 3 with numpy, scipy, numba and moderngl (OpenGL), and ffmpeg; no browser needed
FILES

Download the result

Technical details and provenance
Provider
anthropic
Model
claude-opus-5-5
Agent
Claude Code
Published
Publication
Made by Possible