Peter Yang tests Claude Sonnet 5.5 on seven video projects, from motion graphics written entirely in code to an anime opening assembled with external music and video models. The useful result is less “one prompt makes a film” than a map of when code is enough and when specialized tools help.

What “making video” means here

  • Sonnet is primarily acting as a coding agent: building scenes, animation, transitions, timing, overlays, sound effects, and rendering workflows.
  • The simplest pieces are generated from code alone; higher-fidelity animation uses Suno for music and fal.ai for a video-generation model.
  • Reference videos do much of the creative steering—Yang repeatedly links an example and asks Claude to adapt its style to a new story.

The seven experiments

  1. Motion-graphics reel — a 20-second résumé-style showreel, then revised into a Sonnet launch trailer by changing the copy.
  2. Claude mascot animation — an illustrated walk through human tools, from stone to wheel to AI and AGI, with generated music and effects.
  3. Product-launch film — built with the open-source HyperFrames library, brand assets, preview keyframes, iterative feedback, and music synced to animation.
  4. Vertical short — a website plus a source tweet became a product ad with voice-over; it still needed feedback to remove dead screens.
  5. Talking-head edit — Claude added captions, zooms, overlays, cutouts, and B-roll; technically ambitious, but too visually loud for Yang’s taste.
  6. Animated music video — a mostly one-shot, code-generated piece that was creative and coherent despite simple shapes.
  7. Anime opening — lyrics from Claude, a Suno track, and a video model through fal.ai; more pipeline than pure code, and still below professional animation quality.

The workflow that held up

  • Start from a reference. Give Claude a video whose pacing or visual language you want, then supply your own story.
  • For brand work, use HyperFrames. Feed it fonts, colors, logos, and examples; review keyframes before paying the cost of a complete render.
  • Dictate concrete feedback. Yang adjusted scene duration, visual density, copy, sound, and music over several passes.
  • Separate the modalities. Use a dedicated music model for the track and a dedicated video model when code-generated graphics cannot deliver the needed image quality.
  • Let the agent integrate the pieces. Claude’s strongest role is coordinating assets, writing animation code, syncing to music, and iterating on the assembled result.

What did not work cleanly

  • “Make it creative” can produce visually aggressive, MrBeast-style editing that is capable but wrong for the creator.
  • One-shot claims need qualification: several demos required feedback, source assets, open-source libraries, APIs, or paid generation credits.
  • The anime pieces remain visibly synthetic; Yang says a human editor or animator would still do better.
  • The episode evaluates outputs by inspection, not against a repeatable benchmark or another model on the same seven tasks.

Four practical tips

  1. Link a strong reference instead of trying to describe a style from scratch.
  2. Supply real brand assets and guidelines for product work.
  3. Give Claude a finished music track, then ask it to choose moments and synchronize animation to them.
  4. Add a specialized image/video model only when code-based animation hits its quality ceiling.

“For this one, it was not one shot. I went through a whole process: install HyperFrames, make the keyframes, follow the brand guidelines, give it examples, and then iterate.”