Reading the Rubrics: How Anthropic Evaluated the Teaching Skills

The most quietly radical thing in the k12-teacher-skills repo isn't a skill. It's the evals/ directory — the evaluation framework and rubrics Anthropic used to decide whether the teaching skills were actually good, published alongside a technical write-up of the evals so other education builders can see the method. In a field crowded with tools that ship on a convincing demo and a vibe, publishing the yardstick is the more consequential move. You learn more about a teaching tool from how it was judged than from what it claims.

The headline is simple: the skills were evaluated for rigor, pedagogical alignment, and classroom usability. But the interesting detail is how — the rubrics are broken out by subject and by task, encoded as CSVs, and they include tests for behaviors most AI evaluations never think to measure. Read the rubric list and you get a working definition of what a good K-12 teaching skill has to do. This piece reads it.

Three things the evals were built to measure

The framework judges the skills against three qualities, and it's worth being precise about each because they pull in different directions and a good tool has to satisfy all three at once.

• Rigor — is the content correct and appropriately demanding? A lesson that's wrong, or that quietly lowers the bar, fails here regardless of how polished it looks. Rigor is the guard against confident, plausible, subtly-off output.
• Pedagogical alignment — does the output reflect how the subject is actually taught and how students actually learn it? This is where standards, learning progressions, and subject-specific method matter. A factually correct lesson that ignores the way the concept is built up is still pedagogically misaligned.
• Classroom usability — could a teacher actually use this on Monday? A plan that's rigorous and well-aligned but impractical — wrong length, no student-facing materials, unrealistic for a real period — fails the test that matters most in the room.

Any one of these is easy to optimize alone. The hard part is a skill that's rigorous and pedagogically sound and usable, and the eval framework exists to make sure you can't trade one off silently against the others.

Rubrics by subject, not one rubric for everything

The structure of the evals/ directory is where the design philosophy shows. Rather than a single generic "is this a good lesson?" rubric, the framework carries CSV rubrics broken out by subject, plus two task-specific ones:

The per-subject split mirrors the skills' own references/ directories — the same recognition that a good math lesson and a good ELA lesson are good in different ways, and so must be judged in different ways. A rubric generic enough to cover every subject is generic enough to catch nothing. Encoding them as CSVs, one row per criterion, makes the judgment auditable: you can see exactly what each subject is being held to, and change it if you disagree.

The two rubrics that reveal the mindset

The subject rubrics are expected. The two task rubrics — clarifying_question and differentiation — are the tell, because they measure behaviors that most AI evaluation skips entirely.

The clarifying_question rubric evaluates whether the skill asks good questions before it charges ahead. That's an unusual thing to grade. Most systems reward a fast, complete answer; this framework rewards knowing when you don't have enough to answer well. In a classroom that's essential — a lesson skill that invents the grade level, the standard, or the students' needs rather than asking produces confident, wrong output. Treating "asks the right questions" as a first-class, scored quality is a statement about what safe automation in a classroom looks like.

Grading a skill on the questions it asks — not just the answers it gives — is the difference between a tool that guesses about your class and one that works with you on it. That single rubric captures most of what "trustworthy in a classroom" means.

The differentiation rubric does the parallel job for the tiered-lesson task: it checks that below-, at-, and above-level versions genuinely differentiate and, critically, that the core content stays consistent across tiers. Without that rubric, "differentiation" could pass by simply producing three different worksheets. With it, the skill has to hold the same learning destination across every version — the whole point of principled differentiation.

Rubrics plus real classrooms

Rubrics alone can drift toward what's easy to measure, so the framework wasn't the only check. The skills were refined via early classroom feedback — including at Prospect Schools in Brooklyn — which is how a rubric that looks right on paper gets corrected by contact with a real period, real students, and a real teacher's read of whether the output was usable. The two loops complement each other: rubrics give you consistent, auditable judgment at scale; classroom feedback keeps the rubrics honest about what actually helps.

That pairing is the quietly important lesson for anyone building education tools. Evals without classrooms optimize for the test; classrooms without evals can't be run at scale or reproduced. You need both, and the repo shows both were used.

Why publishing the yardstick matters

There's a reason making the evals public is a bigger deal than making the skills public. Skills tell you what a tool does; evals tell you what "good" was defined as — and in education, where a plausible-but-wrong lesson can quietly cost a class real learning, the definition is where trust lives. A teacher can look at the rubric list and see that rigor, pedagogical alignment, and usability were the bar, and that questioning and differentiation were tested explicitly. A district can point to it. And a builder can fork the framework wholesale.

That last point is the compounding one. Because the rubric CSVs and the framework are open, any education-skill author can fork the evals and evaluate their own tools against a serious, subject-aware standard instead of inventing one from scratch — or shipping on none at all. The per-subject rubrics also pair naturally with the skills' subject reference packs, which encode the same subject-specific thinking on the authoring side. Published rubrics turn one company's quality bar into shared infrastructure for the whole K-12 AI ecosystem.

If you're assembling education tooling now, browse the education agent skills and the wider AI tutoring category for tools you can hold to the same standard, or read more builder insights.

---

Part of the Claude for Teachers series. Related: Fork the Evals · Subject Reference Packs. Browse AI tutoring skills or more builder insights.