The ai-explainer Skill | How Brutalist Builds Explainers

Liam, in for Bear, tears down the skill responsible for the video being watched: ai-explainer. It's the skill invoked for the brutalist Claude cut, cream page, warm ink, one terracotta accent, real narration, and a workflow that reads a skill file end to end before drawing a single frame.

Not a template pack

The easy mistake is calling this skill a template pack. It isn't. It's a workflow that reads a whole skill file, sets the brand and voice, drives every phase, and refuses to publish on its own. A skill here is a folder: a skill file at the top that Claude reads before doing anything, plus the files it references for ai-explainer, the register prose, the brand file, the outro lock, pipeline safety notes, and the parent explainer chassis it extends. The instructions govern; the code just executes them. The pipeline itself runs in a fixed order: scaffold the folder and brand metadata, write narration in the teardown register, generate the voice locally with Kokoro, fill each beat's visual slot from scenes, Manim fragments, or on-screen code blocks, then compile to a 4K master beside the reel. It never uploads and never publishes.

The six-beat spine

The first design decision fixes the shape of every reel: cold open on the Claude composer, a beat where a hesitant writer corrects a real misconception into the reel's actual claim, a body that illustrates the verdict as an artifact card, a handoff that types a paste-ready prompt, and an outro that restates the title. Six moves, the same in every episode.

The interface-only-when-it's-the-subject rule

The second decision is that the Claude interface only appears where the interface itself is the subject: the cold open, the ask, the verdict, and the handoff-outro. Every other beat has to illustrate its concept instead, through a diagram, a Manim fragment, an on-screen code block, or a Remotion illustration. Two composer beats in a row is treated as a smell, and seven videos of composer wallpaper is the specific failure this law exists to prevent.

Ask before receipt

The third decision governs generated visuals specifically: whenever the reel is about to show one, it first shows the Claude prompt that made it, a composer beat with the actual prompt typed in, then the graphic. Never a rendered chart alone, always the ask followed by the receipt. That habit is why the brand feels like the tool it's teaching rather than a slideshow about it.

Where the skill bites, and the Liam substitution

Because the instructions govern everything, an unread paragraph becomes a silent bug: a stale outro, a missing beat, a wallpaper reel. Linting helps, but the discipline stays human. Liam is named as a substitute voice, never a Bear clone; if a real project needs Bear's own voice, that gets rebuilt on the Bear channel specifically, a one-variable change the skill won't fake on its own.

Key takeaways

• ai-explainer is a workflow that reads its full skill file, sets brand and voice, and drives every production phase, not a set of pre-made templates.
• Every reel follows the same six-beat spine: cold open, misconception correction, verdict-as-artifact, illustrated body, paste-ready handoff, and title-restating outro.
• The Claude interface is restricted to beats where the interface is literally the subject; everything else must illustrate its concept some other way.
• Any beat showing a generated visual must first show the Claude prompt that produced it, ask then receipt, never a chart or diagram alone.
• Liam is a named substitute narrator, not a clone of Bear's voice; using Bear's actual voice requires an explicit, separate rebuild.

Who this is for

This is for anyone using Claude Code's brutalist toolkit to build educational explainer videos, and for skill authors who want a concrete example of a markdown file that locks a video's narrative structure, visual rules, and publishing boundaries in one place.