mike-cartoon-broll

Mike Cartoon B-Roll

Create short cartoon inserts that break up talking-head footage. Use one strong opening frame, then give the video model room to invent the progression and scene-appropriate audio.

Style

Use assets/cartoon-style-reference.png as the authoritative visual-style reference in every image-generation request. Treat its characters, objects, poses, and composition as examples, not required scene content. Choose the cast from the requested idea. A scene may include Mike, the robot, both, other characters, or neither. In each frame prompt, list the allowed subjects and say the reference controls rendering only, not scene content.

Warm ivory paper with visible texture.
Fine graphite-and-ink outlines with imperfect cross-hatching.
Muted watercolor and gouache fills.
Sparse editorial compositions with grounded shadows and negative space.
Static wide or medium-wide camera and dry, understated humour.
If Mike appears, use his black cap, charcoal T-shirt, dark jeans, and black sneakers.
If the friendly robot appears, use its rounded weathered teal body, cream face panel, black oval eyes, mustard antenna ball and chest medallion, and segmented dark limbs.
If Mike's dog Sammy appears, draw a chocolate labrador with a dark collar.
The robot stands only for the viewer's coding agent. Never cast it as the database, a function, the computer, or a scheduled job; use objects or other characters for those roles.

Preserve Meaning

Before scene ideation, extract the minimum semantic anchors needed for the point to survive. Do not turn every clause into a requirement:

Must show: only the causally necessary actors, including who initiates the action and who is affected. Treat an explicitly named team practice, workflow, or real-world setting as required when it is part of the claim rather than incidental background.
Must happen: the essential action, who acts on whom, cause, and result.
Must change: any essential before-and-after contrast, role change, or emotional reversal.

Use the full script for context. Treat words such as used to, now, but, because, rather than, and instead as prompts to check for a required change or cause, not as automatic duration rules.

Assume the narration remains audible beneath the B-roll. Let it carry exact wording, lists, setup, and experimental controls. The picture should add the one causal action or reaction that makes the point memorable.

Pick the Scene

How many: a few clips for specific beats, not a cartoon on every section. Mike dropped cartoons across a whole video as "too AI" (Build Your Own Convex, October 2026), then asked for exactly two: a dev asking the title's question at the open, and the "no no no" answer at the verdict. Those worked.

First form the simplest literal scene internally as a meaning check. Then choose one physical-metaphor candidate as well. Pick the clearer, more memorable scene, literal or metaphorical, that preserves every semantic anchor. Humour may compress the narration, but it must not change the actors, agency, cause, result, setting, or required change. The scene need not depict every sentence, but it must preserve the segment's core relationship and change.

Fidelity means preserving the core causal claim, not staging every clause or a controlled experiment. A metaphor may replace technical objects when it still shows the necessary instruction, action, and result. It may transform props inside a required setting, but it must not replace that setting or practice.

Keep the scene sparse. Include only characters and props that serve the idea.
Do not add Mike or the robot merely because they appear in the style reference.
Never turn an externally caused problem into the affected subject causing the problem itself.
A visual joke based on jargon or an idiom is valid when it still shows the real instruction, action, and result. Reject it when the pun replaces the process or required setting.
Do not repeat script wording as generated text or explanatory placards unless the user explicitly asks for it. Prefer visible action and expressive props.
Prefer one decisive physical action over split screens, duplicated comparison subjects, diagrams, or explanatory placards.
Use one continuous action for six seconds, even when it includes a simple cause and result. Use two beats over eight seconds only when two distinct states, roles, or comparisons must be shown separately for the meaning to survive.
Never exceed eight seconds unless the user explicitly asks.

Motion and Continuity

Video models over-animate and lose spatial logic. Write the motion brief to prevent it:

Ask for one brief gesture, then relaxed natural motion. A gesture named in the prompt is otherwise often held for the whole clip (a point that never ends).
Keep animals calm and slow. An animal given a task tends to go frantic.
State depth explicitly for anything that overlaps: "the fence stands in front of the rake", "the kangaroo passes behind the armchair". The model otherwise guesses, and often puts foreground props behind characters.
Do not try to time a gesture to a narration word in the brief. Place the clip so it starts just before the word, then trim or speed it so the gesture peaks on it.
When a clip must match another shot (same room, chair, lighting), make its start frame by image-editing that shot's own frame ("replace the man with a kangaroo"), not by generating the scene again. Separately generated rooms drift in colour, framing, and props, and a cut or dissolve between them jumps.

Workflow

Split the request into one item for every script segment or explicit part the user asked to illustrate. Preserve the exact script lines for each item.
Extract and write the semantic anchors before describing a scene. Record who initiates the action and who is affected when the relationship is directional.
Form a literal baseline and one physical-metaphor candidate. Choose the stronger scene only after checking that it preserves the same actors, agency, cause, result, setting, and change.
For every item, generate the 16:9 raster start frame and draft the loose motion brief in parallel from the same anchors. Use the bundled reference for visual style only. Let the video model choose the exact timing and reactions, and the ending unless the clip must end on a known state (step 8).
Inspect the returned frame at full resolution. Fix identity, anatomy, object-count, fake-text, or composition problems. Reconcile the frame and motion brief against both the anchors and each other. Regenerate the frame if the frame and motion together drop an anchor.
Present one compact approval block for every requested item:
```
Script lines: <exact script segment or explicit user request>
Preserves: <one sentence stating the required relationship or change>
Video filename: <descriptive-slug>.mp4
First frame: <show the generated image>
Motion: <the brief instructions for the video model>
```
Replace the first-frame placeholder with the actual inline image, not only a file path or description.

After all blocks, state the model, resolution, duration, audio setting, cost per clip, and total cost once. Do not bury the approval bundle in extra prose.
Wait for approval. If the user explicitly asks to skip approval, proceed after the same internal frame check.
Generate from each start frame. Supply a last frame only when the clip must end on a known state, such as matching the next shot or leaving an empty set. Make that last frame by image-editing the start frame so the scene matches, and pass it with --last-frame; the script refuses it if the model does not support one.
Review each full result visually at normal speed and inspect a dense filmstrip. Reject identity drift, duplicated props, broken anatomy, camera jumps, style drift, held gestures, frantic motion, or wrong depth order.
Check the generated audio for speech. The model sometimes invents voices, names, or gibberish. Mute a clip with speech; muting is free and matches a silent regeneration. Regenerate with --no-audio only when the take needs redoing anyway, and state the cost. Otherwise preserve the audio for editing.
Never silently pay for a retry. Preserve the failed take and state the retry cost.
When the user flags a problem in a delivered clip, fix that problem first with a trim, speed change, crop, or in-place video edit, and regenerate only if that fails. Ask before replacing a clip the user has already approved; replacements can lose what they liked about the original.
Normalize each generated output to H.264 video at the agreed native resolution and 30fps, preserving the generated audio as AAC. Write it as <VIDEO>/source/generated/<video-filename>.mp4, where <VIDEO> is the video project root and <video-filename> exactly matches the approval block. Create source/generated if it does not exist.
After generation, give the user a clickable Markdown link to the absolute <VIDEO>/source/generated folder so they can open it in Finder and view the videos.

Keep working files together:

generated-video/<clip-slug>/
  first-frame.png
  prompt.txt
  output.source.mp4
  output.source.mp4.receipt.json

Write each output video to:

<VIDEO>/source/generated/<video-filename>.mp4

Never put frames, prompts, raw provider files, or receipts in source/generated.

Video Defaults

Model: bytedance/seedance-2.0
Resolution: native 1080p
Duration: six seconds, or eight seconds for two beats
Aspect ratio: 16:9
Audio: generated
Jobs: sequential unless the user explicitly requests parallel generation

Use scripts/openrouter-video.mjs with --audio to fetch live pricing and cap every paid job. Do not rely on remembered prices. Use --no-audio only when the user requests a silent clip or a take with invented speech is being regenerated anyway (step 10). Use Veo only when the user prefers lower cost or faster generation.

Other OpenRouter video models worth knowing (checked September 2026; confirm live capabilities with the models endpoint before relying on them):

Model	Use for	Notes
`bytedance/seedance-2.5`	Shots that need character or prop consistency	Takes many reference images. 720p maximum on OpenRouter.
`runway/aleph-2`	Fixing one thing in a delivered clip without regenerating it (step 12)	Needs the source as a public HTTPS `video_url`; `scripts/openrouter-video.mjs` does not support video input yet.
`black-forest-labs/flux-video-edit`	Cheap in-place edits to a B-roll clip	Redrew a face in one test, so avoid it on clips with Mike. Outputs 1248x704 at 24fps, so it needs the step 13 normalisation. Do not send `aspect_ratio`.
`kwaivgi/kling-v3.0-pro`	Clips that keep over-animating	Has `negative_prompt` and `cfg_scale` for restraining motion. 720p.

Example motion brief:

Use the supplied frame as the opening composition and visual style. The neat
row of existing dominoes begins falling, but the oversized sticky note catches
the chain at the last moment. Let the action progress naturally and choose a
satisfying understated ending.

Preserve the warm paper, ink-and-watercolor texture, existing subjects, and
existing props. Static camera, one continuous shot, restrained hand-drawn
motion. Generate subtle natural sound effects that match the visible action, but
no speech unless requested. No cuts, unrequested characters, duplicated objects,
generated text, flat vector simplification, 3D rendering, or watermark.

Checklist

The scene earns its place as an attention reset.
Semantic anchors were extracted before scene ideation.
A literal baseline was used to verify the meaning; the final scene preserves every required actor, agency, cause, result, change, and setting.
The affected subject has not been made responsible for an externally caused problem.
The scene communicates through action and props without unnecessary generated text.
The cast and props come from the idea, not from the style reference.
One clear start frame matches the bundled style reference.
The frame is clean when inspected at full resolution.
A frame and draft motion brief were created in parallel for every requested script segment.
Each frame and motion brief were reconciled against the semantic anchors.
Every requested segment has its own approval block with Preserves:.
The motion brief gives direction without choreographing every moment.
The motion brief asks for one brief gesture, calm animals, and explicit depth order where things overlap.
A clip that must match another shot starts from an edit of that shot's frame.
A last frame is supplied only for continuity, made from the start frame.
Live model, duration, resolution, audio setting, and cost are confirmed.
The complete motion is reviewed visually before delivery.
The generated audio was checked for speech, and any clip with speech was muted or regenerated.
The final is H.264 at the agreed native resolution and 30fps with generated audio preserved as AAC.
The output video is saved to <VIDEO>/source/generated/<video-filename>.mp4.
The response links to the absolute source/generated folder.