Comprehension Ladder — Design skill for Claude Code
Claude Code skill that explains anything on Karpathy's comprehension ladder: ASD-STE100 prose (80% strictness), blueprint SVG diagrams, interactive HTML explainers, and narrated 3Blue1Brown-style vide.
How to install Comprehension Ladder
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open bowen31337/comprehension-ladder and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Comprehension Ladder does
Claude Code skill that explains anything on Karpathy's comprehension ladder: ASD-STE100 prose (80% strictness), blueprint SVG diagrams, interactive HTML explainers, and narrated 3Blue1Brown-style videos. Keeps hedges, adds no facts.
Alternatives in Design
- Export Diagram — Export a diagram-design HTML file to .svg and .png next to the source 27.2k ★
- Weather Orchestrator Command — Fetch the current temperature for Dubai, UAE and create a visual SVG weather card 20.2k ★
- Learning Output Style — Interactive learning mode for skill building 14k ★
README
comprehension-ladder
**A Claude Code skill that explains anything so a human can understand it: as Simplified Technical English prose, a blueprint diagram, an interactive HTML page, or a narrated 3Blue1Brown-style video.**
LLMs now do most of the legwork. The human job moves up to oversight and understanding. This skill does the understanding part well. It turns a concept, a diff, a design doc, an incident note or an LLM's wall of text into the format that is easiest to absorb, and it never adds facts or drops hedges along the way.
The idea comes from [Andrej Karpathy's post of 2 October 2026](https://x.com/karpathy/status/2105819303471976479). It describes a ladder of output formats, where each rung is usually easier to absorb than the one below it.
The ladder
| Rung | Output | Good for |
|---|---|---|
| 1. STE prose | Markdown in ASD-STE100 Simplified Technical English, with a strictness setting (default "80% of the way to STE100") | Short explanations, diffs, answers you will re-read |
| 2. Diagram | One self-contained SVG in a blueprint "overview sheet" style: lettered panels, title block, grid reference | Structure, data flow, sequences, incidents |
| 3. Web page | One offline-capable interactive HTML file that runs the real computation | Anything with a parameter, a process over time, or a "what if" |
| 4. Explainer video | Narrated MP4: STE narration + TTS + animation, rendered with Manim or the bundled frames.py renderer |
Topics with a story that changes over time |
Each reply ends with one line that offers the next rung up, e.g. `Next: diagram (2) · HTML (3) · video (4)`.
What makes it different from "just ask the model"
- Same facts on every rung. The skill writes an internal outline first (claims, mechanism, hedges, one example). Every format renders that outline, so climbing the ladder changes the form, never the facts.
- Hedges survive. "May be caused by a race condition" stays
Related Skills
J Explain
Generate a self-contained HTML document or information report that explains a topic — inline CSS/JS/SVG, visua
Aside Visualizer
Aside account-skill port of codexclaw dev-visualizer: visual documents, SVG/HTML explainers and paged PDF repo
Match Resume Style
Regenerate formatting.md and blueprint.md to match the visual style of a reference resume, adapting anything t
Hand Drawn Diagrams
Hand-drawn diagram skill for Claude Code and Codex. Generates monochrome architecture, workflow, and UX bluepr
SVG Sketch Style
SVG Sketch Style Guide — PoC Diagrams
Architecture Diagrams Skill
Diagrams for software and chip design as one interactive offline HTML page: block diagrams, clock trees, pin-l
Related Agents
Diff Explainer
Explains a diff (staged, working tree, branch, or commit) in the project's domain language and then asks the u
YouTube Video Assembler
Turns a finished episode's script.md + visuals/.png + visuals/.mp4 + an already-generated voiceover.wav into a
Comprehension Review
De-warmed cold comprehension-and-structure gate for design.md and track files: reads only the document, never