vibe-knowledge-film: An Agent Skill That Codes Knowledge Videos Frame by Frame
Deterministic code-drawn frames plus a two-layer QC pattern gives developers a reusable alternative to black-box AI video platforms, cutting per-render cost and enabling fine-grained edits. The published prompts and verification workflow are portable to any procedural animation pipeline.
The open-source Agent Skill vibe-knowledge-film turns a text premise into a finished knowledge video without video editing or AI video-platform generation. Instead of generating images, the agent writes code that draws each frame deterministically, so the same frame rendered twice is pixel-identical and any single frame can be edited and re-rendered without touching the rest. The workflow runs through five gates: read the skill's instructions, define the topic and everyday metaphors, build a one-subtitle-per-line storyboard, render sample frames, and export a full video with numeric verification (resolution, duration, file size, presence of audio).
Quality control is split into two layers: a 16-frame collage checks the global composition, and a single frame enlarged to its original size catches subtitles overlapping visuals, labels stacking, and light-text-on-light-background issues. The hands-on test produced two clips, a cold-color piece about folding paper 42 times to reach the Moon (1080p, 30s, ~6MB) and a warm flat-illustration piece about loss aversion (24s), using the same skill with different art styles. All seven prompts are published verbatim, along with three real rework cases and the exact output specs.
Code-drawn frames make the output deterministic, so the same frame renders identically each time and can be edited in isolation, directly attacking the regeneration cost of AI video generators.
The two-layer QC pattern, a collage for global layout and an original-size single frame for details, catches defects thumbnails hide; in this test all issues surfaced only after enlargement.
The three human confirmation gates (topic, storyboard, sample) are what turned a multi-step skill into a reliable pipeline; skipping them would force a full re-render for a single bad frame.