Inside Anthropic’s New Open‑Source Teacher Skills: Architecture, Rules, and How to Use Them

Anthropic, together with Learning Commons, released the k12-teacher-skills repository, exposing the full source code of two official Agent Skills that ship with Claude for Teachers.

Skill Overview

k12-lesson-planning : generates a complete lesson from a single teacher prompt and delivers three Word documents—teacher lesson plan, student materials, and an observation template.

k12-lesson-differentiation : takes an existing lesson and produces three tiered versions (below, at, above grade level) plus a teacher‑level differentiation plan, resulting in four Word documents.

Both skills connect to the Learning Commons Knowledge Graph, automatically aligning content with U.S. standards such as CCSS, NGSS, TEKS, and SOL.

Lesson‑Planning Workflow

Route (silent routing) : detect the subject (math, ELA, science, social studies) and load the corresponding references/{subject}.md file; failure to load is treated as a critical error.

Clarify : ask up to two key clarification questions, using defaults for everything else.

Ground : if the Learning Commons KG is available, call find_standard_statement to fetch the exact standard text.

Draft offer : let the teacher choose between a full output or a draft preview; the default is the full version.

Build : construct the lesson structure according to the subject‑specific template.

Output : write a lesson.json file and run a rendering script that produces the Word documents.

The final deliverables are: lesson_plan.docx: teacher’s lesson plan with phases, timing, script, and preset error handling. student_materials.docx: printable student worksheets (generated only when needed). observation_template.docx: observation record template with look‑fors, misconception table, and exit‑ticket buckets.

Lesson‑Differentiation Workflow

Based on the U.S. practice of Tiered Instruction, the skill receives an existing lesson (uploaded, pasted, linked, or referenced) and outputs: teacher_plan.docx: differentiated plan with entry points and flexible‑grouping criteria for each tier. worksheet_group_a.docx: "Below" tier worksheet. worksheet_group_b.docx: "At" tier worksheet. worksheet_group_c.docx: "Above" tier worksheet.

The three versions share the same core tasks and materials; only scaffold density and extension questions differ, following the Tomlinson differentiation framework.

Eight hard rules (R1‑R8) govern the output, e.g.:

R6 – Invisible modifications : student worksheets never reveal their tier; group names A/B/C are used instead of Scaffolded/Guided/Independent.

R7 – Progressive scaffolding : tasks become lighter across the three steps, encouraging independence.

R4 – Below‑tier scaffold cap : limits scaffold density to avoid over‑reliance.

R8 – UDL defaults : when no learning‑disability information is provided, all tiers receive sentence supports and a vocabulary list.

Design Pattern 1: JSON‑Driven Content Separation

The skill never lets the LLM write Word directly. Instead, the LLM produces a lesson.json file, which a bash script renders into the final documents:

LLM writes lesson.json (content source)
    ↓
bash scripts/render_all.sh lesson.json $OUTPUT_DIR
    ↓
Word documents (rendered output)

lesson.json

contains two top‑level fields: shared: a registry of reusable items (questions, standard excerpts, vocab, misconception tables) referenced by a key such as t1 or prices_table. documents[]: an array where each document references shared items via {"type": "from_shared", "key": "t1"}.

This design lets a teacher edit a single shared entry and have all four output documents update automatically, guarantees identical core task text across tiers, and ensures consistent formatting across Word, HTML, and JSON.

Design Pattern 2: On‑Demand Subject References

Each subject has its own references/ markdown file (e.g., math.md, science.md) containing dozens of kilobytes of discipline‑specific teaching guidelines, non‑negotiable options, and a mapping to lesson.json. The skill’s main file simply routes to the appropriate reference file, keeping the LLM’s context short while allowing independent evolution of each subject’s knowledge base.

Design Pattern 3: Density Rules & Language Guardrails

Paragraphs or labeled blocks are limited to three sentences; longer text must be split, listed, or placed in a table.

List items are fragments: one idea, ≤15 words, never joined with semicolons.

Tiered content must be presented in a single table (rows = phases, columns = tiers, ≤25 words per cell).

Continuous prose exceeding half a page must be restructured into tables or separate sections.

Standard excerpts appear verbatim only inside a target‑standard callout; elsewhere they are summarized in ≤10 words.

A copyright guardrail prohibits verbatim copying of existing curricula (e.g., Illustrative Mathematics) and forces original phrasing.

Integration Layer

The repository’s .mcp.json registers nine education‑SaaS micro‑service connectors (ASSISTments, Brisk Teaching, Canva, Coteach, Diffit, Eedi, MagicSchool, Snorkl, TeachFX). Combined with the Learning Commons Knowledge Graph, Claude can retrieve:

Standard text for all 50 U.S. states.

Prerequisite skills and learning pathways for each standard.

Concrete learning components linked to each standard.

This makes the skill a full‑stack AI teaching assistant capable of replacing a half‑staff lesson‑planning team.

How to Use

Claude for Teachers users have the skills pre‑installed. For other users, install via Claude Code:

git clone github.com/anthropics/k12-teacher-skills
claude plugin marketplace add ./k12-teacher-skills/plugin
claude plugin install k12-teacher-skills@k12-teacher-skills

After installation, a single natural‑language prompt triggers the appropriate skill, e.g.:

"I'm teaching long division to 5th graders" → lesson‑planning.

"Need to teach photosynthesis tomorrow" → lesson‑planning.

"My students are at different levels" → lesson‑differentiation.

"Adapt this lesson for below and above grade" → lesson‑differentiation.

The skill’s description defines trigger conditions in legal‑style language and explicitly forbids invoking the differentiation skill when a new lesson already requests tiered materials.

Applicability to Chinese Education

Core engineering patterns (JSON‑driven separation, on‑demand references, density rules, teacher‑student language split) are language‑agnostic and can be reused.

Content layers must be rewritten to match Chinese standards (e.g., the 2022 national curriculum) and replace the Learning Commons KG with a domestic knowledge base.

A local MCP ecosystem (equivalents of MagicSchool, Brisk Teaching, etc.) is required for full functionality.

Evaluation Framework

An evals/ directory provides a complete testing harness for the skills, demonstrating how to assess correctness, adherence to hard rules, and output quality.

Conclusion

Anthropic’s open‑source repo serves two purposes: a production‑grade AI lesson‑planning tool for U.S. teachers and a comprehensive tutorial for skill developers. It showcases how to handle vertical knowledge, keep multi‑document consistency, enforce output quality, define precise trigger conditions, integrate third‑party MCPs, and evaluate the result. The author rates the repository 95/100, noting that the dense SKILL.md documentation raises the initial learning curve but every line is practical engineering insight.