K TeachField Manual
GitHub
Open Source · MITPlatform-NeutralDeterministic CLINode ≥ 22.18

Turn any AI agent
into a persistent, mission-driven teacher

K Teach is an open-source skill and deterministic CLI that follows the open Agent Skills format. It uses a structured teaching loop, a single-source Lesson Bundle mechanism, and cross-session evidence tracking to make AI truly teach you — not just answer you.

Manual Excerpt — 01 · Initializebash
$ npm install -g k-teach@latest
✓ installed k-teach 1.x

$ k-teach init --tools codex,claude --teach mathematics
→ created .k-teach/config.yaml
→ created teachs/mathematics (mission, glossary, logs)
→ wired Agent integrations for codex, claude
✓ Learning Project ready. Start teaching.
3
Channel outputs: Web · PPT · WeChat
Mission & evidence across sessions
1
Single-source Lesson Bundle
0
No secrets stored · private by default
§01 Core Features

Not just answers,
but teaching

K Teach breaks teaching into verifiable, traceable, reproducible steps, so the AI teacher has a mission, a rhythm, and evidence — like a human teacher.

F.01

Mission-driven teaching

Establish an observable learning mission and success criteria first, then design each "smallest useful challenge" around it.

F.02

Single-source Lesson Bundle

Explanations, exercises, feedback, and retrieval rules are stored as one authoritative source; channel outputs are only derived from it.

F.03

Evidence-driven learning

We only record "demonstrated learning" when the learner explains, retrieves, applies, transfers, or corrects.

F.04

Persistent across sessions

The learning workspace keeps the mission, content, evidence, and teaching decisions across sessions — pick up where you left off.

F.05

Strict safety boundaries

No credentials in storage, local lessons private by default, public publishing requires confirmation, artifacts immutable after release.

F.06

Platform-neutral

SKILL.md is a platform-neutral contract shared by Codex / ChatGPT / Claude / any agent.

§02 Teaching Loop

A complete, repeatable teaching loop

From mission to evidence, an eight-step tight loop. Each step is verifiable by the CLI and traceable to the Lesson Bundle.

01

Set the learning mission

Define concrete goals and observable success criteria so teaching has direction.

Mission
02

Assess prior knowledge & preferences

Check existing knowledge, learning records, trusted sources, and study habits.

Assess
03

Pick the smallest useful challenge

Choose a task slightly above the current level that you can start right away.

Challenge
04

Verify facts against trusted sources

Validate every factual claim against authoritative primary sources — no fabrication.

Verify
05

Build the Lesson Bundle

Focus on a single lesson's capability: activate prior knowledge → explain → practice → feedback → retrieval without looking back.

Bundle
06

Choose teaching media

Text, deterministic SVG diagrams, or optional generated visuals — combine as needed.

Media
07

Resolve intent & generate

When intent is unclear, ask first: learning / WeChat / PPT — each derived independently.

Generate
08

Record demonstrated learning

Write evidence back to the learning record to drive the next round of content choices — the loop continues.

Evidence
§03 Multi-channel Output

One course, three outputs

Web lessons, HTML presentations, and WeChat articles are all deterministically derived from the same Lesson Bundle — consistent facts, traceable sources.

WEB

Web Lessons

A local static learning site; every lesson opens independently, with exercises and retrieval checks.

  • Complete learning experience, offline-ready
  • 7 switchable Teaching Themes
  • paper / night / print display modes
PPT

HTML Presentation

Derived from a Presentation Brief, with keyboard navigation and presenter notes.

  • Classroom teaching / talk dual use
  • Keyboard navigation + presenter notes
  • Shares visual themes with Web
WX

WeChat Article

Derived from a Publication Brief, requires explicit authorization, platform-safe expression.

  • Independent Channel Theme
  • Draft / preview / publish state machine
  • No exercises, stays platform-compliant
§04 Tech Stack

Built for reproducibility

TypeScript source, deterministic CLI, open Agent Skills format — stable, verifiable, easy to publish.

TypeScript
Node.js ≥ 22.18
pnpm
Vite preview
Open Agent Skills
{} JSON CLI
📦 npm distribution
🔁 Changesets
§05 Quick Start

Three commands to start a learning project

Install the persistent CLI globally, initialize a Learning Project, and let AI take over the teaching.

bash
# 1. Install the persistent CLI globally
npm install -g k-teach@latest

# 2. Enter your project and initialize
cd your-learning-project
k-teach init

# 3. Specify agents and first Teach (optional)
k-teach init --tools codex,claude --teach mathematics

# Local preview (binds 127.0.0.1 only)
k-teach preview --open
Install CLI
Install k-teach globally to get deterministic commands and preview.
Initialize project
Generate .k-teach/config.yaml, the teachs/ collection, and an initial Teach.
Connect agents
Generate per-platform entries from the same authoritative Skill.
Start teaching
Tell the agent: "Use K Teach to help me learn distributed systems."