Five Prompts, Zero Code: Running a Stick-Figure Video Skill in WorkBuddy to Three MP4s
Turning AI video generation into a repeatable, checkable pipeline means the outcome stops being a lottery: the planning comes from the model, but the pixels come from deterministic local code, and the character stays stable because the drawing rules are literal. The character-anchor trick also shows a cheap, effective way to fight AI inconsistency without heavier tooling.
WorkBuddy's stickman-video-director skill is not a single mega-prompt. It installs as a folder: SKILL.md plus 52 contract files in references/ and 41 scripts in scripts/. Each contract owns one concern, from style catalog to voice continuity, storyboard templates, and acceptance checklists. In director mode (the A mode tested here), a single sentence produces a full proposal: hook, five-beat emotional arc, three accent-color meanings, BGM direction, and an English narration split into 10-second blocks at 20–25 words per block.
The most notable piece is the character anchor. Instead of writing 'keep the character consistent', the skill emits a frozen specification: hollow circular head, no face, no eyes, black line art, exactly two arms and two legs, and a rule that raising an arm replaces the previous line, never leaving a resting arm underneath. The author reports this is precisely to stop the model from drawing an extra hand. Each of the three clip prompts is self-contained with 12 sections, so any of them can be dropped into any prompt-to-video tool.
Rendering happens locally via a Python script (PIL per-frame drawing plus an ffmpeg pipe) and purely on CPU. The outputs: a 30-second 1280×720 MP4 at 1.1MB and a 10-second 720×1280 MP4 at 0.43MB, both verified with ffprobe for H.264/AAC and exact durations. Switching to a second style (Style 2A) means swapping the entire character anchor and color discipline, not just renaming a style.
The character anchor treats consistency as an engineering spec instead of a model behavior to hope for; that's a generalizable pattern for any AI generation where identity must persist across segments.
The skill's structure as a set of single-purpose contracts plus acceptance checklists is what makes it repeatable; the model isn't being asked to remember the whole production, it's following a process with explicit gates.
Dividing the work into model-authored prompts and deterministic local rendering means the only AI-indeterminate part is the planning; the actual pixels are fixed, so the output is verifiable and editable.
The article shows a full workflow, but it also exposes a boundary: full auto mode (B) requires npm deps and a heavier environment; the author prefers reusing the prompt output with a local script, suggesting the skill's value is in the planning layer, not the renderer.