⏳ This skill is pending AI review.
Scores will appear once the review pipeline completes.
legible-explainer
Explain a subject in the format that is easiest to understand, following Karpathy's output-format ladder: ASD-STE100 Simplified Technical English prose, then a diagram, then a self-contained HTML spec sheet drawn like an engineering drawing, then a narrated explainer video. Use it when the user asks for an explanation as the deliverable: explain a concept, protocol, system, codebase, spec, plan, diff or a long agent-written output; help them understand text they cannot follow or have no time to read; or asks for STE, ASD-STE100, Simplified Technical English, controlled language, 'explain it 80% STE', a diagram to keep open while they work, an explainer page or one-page overview, a cheat sheet or spec sheet, an explanation 'in HTML', or a 3Blue1Brown-style explainer video. Not for production UI, landing pages, dashboards, slide decks, promo videos or diagrams the user wants to edit.
Choose how to use this skill
You do not need every option. Choose the path your AI client supports. The stable page stays the same; versioned files are immutable.
1. Native installer
This listing has no registered native installer command. Use the complete package or source fallback below, depending on what your client supports.
Do not guess an installer command or replace an existing version without reviewing the diff.
2. Complete package recommended
Download the ZIP when available. It includes SKILL.md plus the references, security notes and version metadata.
No complete ProSkills package is published for this listing yet.3. Prompt-only
Copy the prompt above when the agent can read the stable page or when you want to adopt the workflow without installing a skill.
Need only the instruction file?
Download SKILL.md only if your client requires a single file. The complete ZIP is safer for a full installation because it preserves the references and release context.
No path installs or executes anything by itself. Your agent still needs access to the project files. Before updating, compare the installed version and review the diff.
// RATINGS
Not yet listed on ClawHub or SkillsMP
// README
The Karpathy Ladder
An unofficial Claude Code skill that makes LLM output easy to understand. Not affiliated with or endorsed by Andrej Karpathy or Anthropic. Ask for an explanation and get it in the format that fits: ASD-STE100 prose, a diagram, a one-page engineering-drawing spec sheet, or a narrated 3Blue1Brown-style video.
On 2 October 2026 Andrej Karpathy posted that we will spend a lot more time trying to understand the outputs of language models, and he ranked the output formats that help most. Each one is "even better" than the last: text in ASD-STE100 (the controlled English of aircraft maintenance manuals), then diagrams, then web pages, then custom explainer videos. This repo turns that ladder into one skill that picks the right rung for each question, builds the output, and checks it before you see it.
Install
In Claude Code (v2.1.275 or later):
/plugin install karpathy-ladder --marketplace FutureAtoms/karpathy-ladder
On an older version, add the marketplace first:
/plugin marketplace add FutureAtoms/karpathy-ladder
/plugin install karpathy-ladder@karpathy-ladder
The plugin is one skill. It has no hooks and no MCP server; it runs two local Python checkers on the pages it makes.
Or copy the skill folder by hand:
git clone https://github.com/FutureAtoms/karpathy-ladder
cp -r karpathy-ladder/skills/legible-explainer ~/.claude/skills/
The page checker needs Python 3 and Playwright (pip install playwright && playwright install chromium). The video rung also uses Manim Community Edition, ffmpeg, a LaTeX install, and kokoro-onnx for a free local voice (or ElevenLabs when ELEVENLABS_API_KEY is set).
Use it
Ask the way you normally would. The skill picks the lowest rung that fully answers the question, and an explicit format request always wins.
| You ask | You get |
|---|---|
| "Explain TCP congestion control in ASD-STE100, about 80% of the way" | The answer in chat, in STE: short sentences, approved verb forms, one meaning for each word |
| "Draw the OAuth PKCE flow so I can keep it open while I debug" | A self-contained HTML page with a hand-drawn SVG sequence diagram, every value on its arrow |
| "Make a one-page overview of X I can pin next to my desk" | An engineering-drawing spec sheet in HTML, plus A4 and A3 PDFs |
| "An agent wrote this plan overnight. What does it propose and what could go wrong?" | A sheet that maps the document, checks it against itself, and leads with a verdict you can say at standup |
| "Make a 3b1b-style video on self-attention with a free local voice" | A narrated MP4 built with Manim and Kokoro: a title card that says what is coming, the explanation at a calm pace, a recap card, burned-in captions and a quiet music bed |
What is inside
skills/legible-explainer/
SKILL.md how to pick a rung, write in STE, deliver
references/ste-writing.md ASD-STE100 rules in plain words, the 80% / strict dial, a substitution table
references/diagram.md rung 2: SVG diagrams
references/spec-sheet.md rung 3: the drawing-sheet layout and its components
references/video.md rung 4: Manim or HyperFrames, voice, narration, scene plan
assets/spec-sheet-template.html a working sheet with every component
scripts/ste_lint.py STE linter: sentence limits, verb forms, passive voice, unapproved words
scripts/check_page.py renders at 5 widths and in print; flags overflow, collisions, tiny print type
evals/ the test prompts used below
examples/ real outputs from the evaluation runs
Every page the skill makes opens from file:// with no build step, and Google Fonts is its only network request. The skill also follows a house style that you can edit in SKILL.md: no em dashes, justified body text, no one-sided accent bars, no text squeezed into a narrow column.
Does it work?
The skill was built with Anthropic's skill-creator loop and reviewed by four AI reviewers from three model families. As of 2026-10-03:
- Blind reader test: each reviewer saw the skill's output and plain Claude's output for six tasks, labelled A and B at random, and said which one a reader would understand faster. The skill won 17 of 24 comparisons, with 2 ties.
- Graded checks: 93% passed with the skill, 75% without it, across the six evals (STE quality, accuracy, layout, print, delivery).
- Cost: about 390 s more per task than plain Claude, mostly spent on checking the page and verifying facts.
Both sides ran fresh on the same day with the same model. One run per side is a small sample, so read these as a direction. Details, the tasks where plain Claude won, and what changed after are in docs/benchmark.md.
Examples
| Rung | Example |
|---|---|
| STE prose | TCP congestion control, mutex vs semaphore |
| Diagram | OAuth 2.0 + PKCE with Next.js and Auth0 |
| Spec sheet | ASD-STE100 overview (HTML, A4 and A3 PDF), review of an agent-written migration plan |
| Video | Self-attention in 92 s (MP4, script, scene plan, Manim source) |
Open an HTML example in a browser. The sheets reflow on a phone and print on one landscape page.
Contributing
New examples, sharper checks, more STE substitutions and new video styles are all welcome. Start with CONTRIBUTING.md, pick an issue labelled good first issue, or post what the skill made for you in Show and tell. Run ./tests/check.sh before you open a pull request; CI runs the same checks.
Credits
The idea and the ladder come from Andrej Karpathy's post. ASD-STE100 is maintained by the ASD Simplified Technical English Maintenance Group; the standard is free from asd-ste100.org. This skill paraphrases its rules and does not reproduce the standard.
MIT licensed. Made by FutureAtoms.
// HOW IT'S BUILT
KEY FILES