Report To Human — AI skill for Claude Code
让 AI 说人话 — a 4-layer agent-to-human reporting protocol inspired by Karpathy's output ladder: ASD-STE100 @80% prose → iconified settings → HTML panel → thin chat entry.
How to install Report To Human
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open guoshaoyang-pku/report-to-human and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Report To Human does
让 AI 说人话 — a 4-layer agent-to-human reporting protocol inspired by Karpathy's output ladder: ASD-STE100 @80% prose → iconified settings → HTML panel → thin chat entry. Agent skill for Claude Code / Codex / Verdent.
Alternatives in AI
- Upstream Watch — Review pydantic-ai and pydantic-ai-harness releases published since the last watch entry, and record a verdict 361 ★
- Shuorenhua — 说人话|Chinese-first AI writing refinement skill that reduces boilerplate and AI tone while preserving facts, ter 262 ★
- Sloptrim — A local detector for AI-writing patterns 196 ★
README
report-to-human
**让你的 Agent 学会 REPORT TO HUMAN。让 AI 说人话。**
A 4-layer reporting protocol for AI agents, inspired by [Karpathy's output ladder](https://x.com/karpathy/status/2105819303471976479) (2026-10-02): writing → diagrams → HTML → video.
Your agent should not dump three screens of markdown into the chat box. The chat box is the thinnest layer. The content lives in the layers above it.
| Layer | Carries | Rule |
|---|---|---|
| L1 Prose | reports & analysis | ASD-STE100 at 80% strictness — the controlled language built for aircraft maintenance manuals. Short sentences, one name per concept, active voice, every number and qualifier preserved |
| L2 Icons & diagrams | settings & structure | config as a one-line icon badge row (🤖 model · 🧮 algorithm · 📈 lr · ✂️ cap · 🖥️ cluster), structure as diagrams, curves as figures |
| L3 HTML panel | the main reading surface | one self-contained interactive HTML file: settings chips, main figures, conclusions, sample browser |
| L4 Chat box | entry point only | name + one-line takeaway + panel link + report link. Nothing else |
flowchart TB
subgraph content["where the content lives"]
direction LR
L1["L1 · Prose\nASD-STE100 @ 80%"]
L2["L2 · Icons & diagrams\nsettings · structure · curves"]
L3["L3 · HTML panel\nmain reading surface"]
end
L4["L4 · Chat box\none-line takeaway + links"] --> content
Why ASD-STE100?
ASD-STE100 is the controlled language the aerospace industry built so a mechanic anywhere in the world reads a step once and gets it right: ~900 approved words, one meaning each, short sentences, active voice, one instruction per sentence. Karpathy's tip: ask for **"80% of the way to ASD-STE100"** — keep the discipline, relax the vocabulary lockdown. It fixes the exact things AI prose breaks: synonym rotation (worker → agent → executor), hedge stacks (should/might/could), 60-word sentences, and decorative filler.
Full rul
Related Skills
ASD SDE 80 Karpathy
Agent skill that makes AI answers clear: short sentences, plain words, one idea per sentence (80% ASD-STE100).
Plain Technical Writing Skill
Claude Skill: plain technical writing an expert reader can find, understand, and use — ISO 24495-1 reader mode
Recovery Classification
Enforces Law 4 (Verify Before Reporting) of the 7 Laws of AI Agent Discipline. After any failure in the verifi
Xiaoyaoclaw Usage Report
OpenClaw usage & performance reporting - parse session JSONL for task duration, tool/skill/model usage, token
Strip AI Tells
Use when prose a human reads as writing sounds machine-written and needs an edit that fixes it rather than a r
LLM Wiki Skills
Turn your markdown vault into a compounding knowledge wiki (Karpathy inspired). Six agent skills - knowledge g
Related Agents
Browser Scout Sonnet
Default browser tier. Read-only work in the user's real Chrome — open a ticket, a chat channel, a dashboard or
Codex Adversarial
Thin shim around codex exec review — runs codex directly, writes audit entry, returns terse verdict+count. Use
Prep Librarian
Final stage of the LectureLoop pipeline. Assembles the per-subject results into Output/ /INDEX.md and produces