understanding-ladder

Agent skills that make LLM explanations easier to understand — built on Karpathy's understanding ladder.

English | 中文

npx skills add ChangWenC/understanding-ladder

Live demo → one explanation, rewritten at every rung.

Why

In October 2026, Andrej Karpathy wrote that LLM output is nearly free; understanding it is the bottleneck. So stop accepting walls of text and ask for formats that are easier to understand. He ranked four, each "even better" than the last:

The hard part is not knowing the ladder. It is picking the right rung every time without being asked — and not climbing higher than the question needs. This repo turns that into a skill.

What you get

understanding-ladder — which format

Classifies the content, recommends the lowest rung that fully answers, and produces it:

Three rules hold on every rung:

• Nothing cut. Simple language, not less content. Every number, condition, and hedge survives.
• Draft first. Every diagram, page, or video starts from a rung-1 text draft you can check.
• "What to check" at the end. Easier to understand ≠ verified.

Works in English and Chinese — it replies in your language.

→ SKILL.md · English examples · Chinese examples

plain-chinese — rung 1 for Chinese

In English you can just say "in ASD-STE100" and the model knows the rules. Chinese has no such name. plain-chinese spells the rules out: 15 of them, grouped into words, sentences, paragraphs, what to delete, and what to keep. understanding-ladder calls it automatically for Chinese output.

→ SKILL.md · 5 before/after pairs

Demo: one explanation, every rung

Topic: why does server latency explode once utilization passes ~80%?

Open the live demo (English / 中文 toggle, single file, no dependencies). Tab 3 has a slider: drag utilization from 90% to 95% — five more points, double the latency.

Install

Any agent that supports Agent Skills (Claude Code, Codex, Cursor, …), via skills:

npx skills add ChangWenC/understanding-ladder

Claude Code plugin:

/plugin marketplace add ChangWenC/understanding-ladder
/plugin install understanding-ladder@understanding-ladder

Manual (Claude Code global skills directory):

git clone --depth 1 https://github.com/ChangWenC/understanding-ladder.git
cp -r understanding-ladder/skills/* ~/.claude/skills/

Use

Just ask. The skill triggers from its description:

Help me understand how learning rate affects convergence.
Explain the TLS 1.3 handshake the clearest way.
An agent wrote this plan — help me see what it actually does.
Walk me through the Fourier series from rung 1 to rung 4.

No install? Paste this (Karpathy's "80%" plus the one rule that matters most):

Write about 80% of the way to ASD-STE100: one idea per sentence, one term per meaning, active voice, no filler. Cut nothing — keep every fact, number, condition, and hedge; just make the sentences clean.

Limits

• Easier to understand ≠ verified. The ladder lowers the cost of understanding, not of checking. A polished artifact lowers your guard; an error in an animation is more convincing than in text. Hence "What to check" on every output.
• Rung 4 writes the storyboard; rendering is done by a video tool. showtime 0.4.0+ reads the table directly (showtime new <template> <dir> --from-storyboard storyboard.md, English or Chinese headers).
• plain-chinese is a lightweight rule set for LLMs, not a full writing standard. Sentence-length numbers are rules of thumb, not tested in reading studies.
• The examples were produced by following the skills and checked by hand. They show direction, not a benchmark.

Related work

Ideas only; no content copied.

Credits

• The ladder comes from Andrej Karpathy's post.
• Rung 1 is inspired by ASD-STE100; not affiliated with ASD. No STE rule text or dictionary is reproduced. ASD-STE100 is a registered trademark of ASD.
• Rung 4 tools: showtime, Manim, Kokoro.

License

MIT