Inside k12-lesson-differentiation: Tiered Lessons Below/At/Above Level
A close read of the second open-source Claude for Teachers skill — how k12-lesson-differentiation turns one lesson into below-, at-, and above-level versions while keeping the core content identical, and the rubrics that keep it honest.
The second skill Anthropic open-sourced with Claude for Teachers tackles the task teachers name most often when you ask what eats their evenings: taking one good lesson and making it work for a room where students are in three different places. k12-lesson-differentiation is the skill for exactly that, and like its sibling it lives in the open at github.com/anthropics/k12-teacher-skills. Reading it teaches you what real, defensible differentiation looks like when a machine does the first draft.
The one-line description undersells the discipline in the design. The skill adapts an existing lesson into tiered versions — below, at, and above proficiency level — and for specific student needs, while keeping the core content consistent across every tier. Its output is a differentiation plan plus personalized, student-facing materials for each proficiency level. The constraint doing the heavy lifting there is "core content consistent." That's the line between differentiation and just handing weaker students a different, easier lesson — and the repo shows how the skill holds that line.
Like its sibling, this skill was co-developed with Learning Commons, designed around tasks teachers said mattered most, and refined through early classroom feedback, including at Prospect Schools in Brooklyn. Inside Claude for Teachers it powers the "differentiate for every learner" use case — you hand over a lesson and get back a differentiation plan plus per-proficiency student materials. What follows is a look under that hood at how the skill produces tiers a teacher can actually defend.
Differentiation, not dilution
Bad differentiation quietly changes the destination. A below-level student gets a watered-down topic, an above-level student gets an unrelated enrichment worksheet, and by Friday three groups have learned three different things. Good differentiation changes the route, not the destination: everyone is heading for the same standard and the same core content, with the scaffolding, pacing, and challenge adjusted along the way.
k12-lesson-differentiation is built around that principle. It starts from a lesson you already have and produces tiered versions that share a common core, so the below-level tier gets more scaffolding toward the same idea and the above-level tier gets more challenge with the same idea — not a detour. Because the skill is open, you can verify this is a design commitment and not marketing language. The instruction to keep core content consistent across tiers is baked into how the skill operates, which is why the versions it produces stay pointed at one learning goal.
Here's how the three tiers relate under that constraint:
| Tier | What changes | What stays the same |
|---|---|---|
| Below proficiency | More scaffolding, smaller steps, additional support toward the concept | The core content and the target standard |
| At proficiency | The lesson as designed, on grade level | The core content and the target standard |
| Above proficiency | Added challenge, extension, deeper application | The core content and the target standard |
That right-hand column is the whole point. Adjust the left column freely; never touch the right.
Two rubrics that keep it honest
The most instructive part of the differentiation skill isn't the skill — it's how Anthropic evaluated it. In the repo's evals/ directory, alongside the per-subject rubrics, are two that speak directly to this task: a differentiation rubric and a clarifying-question rubric. Together they define what "good" means for this skill in a way you can inspect.
The differentiation rubric is the obvious one: it judges whether the tiered output actually differentiates well — appropriate scaffolds below level, genuine challenge above, and that all-important consistent core across the tiers. The clarifying-question rubric is the subtler and more interesting one. Differentiation is impossible without knowing the room. A skill that charges ahead and invents student needs is worse than useless — it produces confident, wrong tiers. So the skill is evaluated in part on whether it asks the right clarifying questions before differentiating: what proficiency levels are present, what specific student needs to account for, what the existing lesson already assumes.
A teaching skill that never asks a question is a liability. Evaluating for good clarifying questions — not just good final output — is how you build a differentiation tool a teacher can actually trust with a real class.
That evaluation choice reveals the mindset. The team treated "does it know what it doesn't know" as a first-class quality, and wrote a rubric for it. If you're building your own education skills, that's the detail to copy.
Subject knowledge and standards come from files, not vibes
Like the lesson-planning skill, differentiation ships with per-subject references/ — ela.md, math.md, science.md, social_studies.md — plus learning-commons-kg.md. This matters more than it might seem. Differentiating an ELA close-reading lesson is a different craft from differentiating a math problem set; the scaffolds that help below-level readers are not the scaffolds that help below-level problem-solvers. Per-subject reference files let the skill differentiate the way an experienced teacher of that subject would.
The learning-commons-kg.md reference connects the skill to the Learning Commons Knowledge Graph, which supplies standards across all 50 states and the learning progressions beneath each one. Those progressions are what make principled tiering possible: if you know the sequence of competencies a standard is built from, you know what "below level" concretely means for this content — which earlier competency to scaffold back to — and what a real "above level" extension is. The tiers aren't guesses; they're placed along a known progression.
Try it, fork it, or adapt it
Because it's open source, you're not limited to using the skill inside Claude for Teachers. Inside the product, it's already installed and wired to the Learning Commons connector — you hand it a lesson and ask for tiered versions. Outside, you can install the plugin in Claude Code:
git clone https://github.com/anthropics/k12-teacher-skills
claude plugin marketplace add ./k12-teacher-skills/plugin
claude plugin install k12-teacher-skills@k12-teacher-skills
The skill assumes the Learning Commons connector; without it, you supply your own standards and curriculum context and the skill still tiers and renders. And if you're an education builder, the two rubrics in evals/ are a gift — a working definition of what a differentiation skill and its questioning behavior should be judged on, ready to fork for your own tools.
The through-line with its sibling skill is clear. Lesson planning gets you a strong first lesson; differentiation makes that lesson reach the whole room. Both keep the human in the judgment seat and put the grunt work — the drafting, the tiering, the rendering — on Claude. If you want the pedagogical case for why this task is the one worth automating first, the companion piece on one lesson at three readiness levels makes it. For hands-on marketplace tools in the same spirit, see lesson-plan-studio and the wider education agent skills, or browse the full AI tutoring category.
Part of the Claude for Teachers series. Related: One Lesson, Three Readiness Levels · Inside k12-lesson-planning. Browse AI tutoring skills or more builder insights.