Karpathy just explained how we’ll work with AI next: diagrams and HTML pages instead of walls of text.
Someone already packed it into one skill.
It is called visual-explainer. Your agent turns plans, diffs and recaps into self-contained HTML pages or slide decks that open in the browser.
Repo: /nicobailon/visual-explainer
Install in Claude Code:
/plugin marketplace add nicobailon/visual-explainer
/plugin install visual-explainer@visual-explainer-marketplace
Some commands worth trying first:
1. /diff-review: a visual code review with an architecture comparison
2. /plan-review: checks a plan against your codebase, with evidence for every fix
3. /project-recap: a mental model of the project when you come back after a week
4. /fact-check: verifies a doc against the actual code
And when the agent is about to print a table with 4+ rows, it renders a page instead.
10.5k stars, MIT, works with Claude Code, Codex, Cursor and Pi.
Anthropic engineers already learn their own codebase this way. They ask Claude for a visual explainer of the slice they are about to change.