Subject-Specific Rigor: The ELA, Math, Science, and Social-Studies Reference Packs

Open either of the teaching skills in the k12-teacher-skills repo and you'll find the same telling detail: a references/ directory holding one file per subject — ela.md, math.md, science.md, social_studies.md — plus learning-commons-kg.md. That structure is easy to skim past, but it encodes the single most important design decision in the whole thing. The skills refuse to treat "teaching" as one uniform task. They treat English language arts, mathematics, science, and social studies as four different crafts that happen to share a building.

This matters because the default failure mode of AI lesson tools is exactly the opposite. Feed a generic model "write a lesson on X" and it reaches for a generic lesson shape — objective, warm-up, activity, exit ticket — and pours any subject into it. The result is superficially fine and pedagogically flat, because a good science lesson and a good ELA lesson are not the same object wearing different content. The per-subject reference packs are how Claude for Teachers avoids that trap. This piece is about what those files buy you and why one-size-fits-all is the wrong size for every classroom.

What lives in a reference file

A reference file is domain grounding: the subject-specific guidance the skill consults when it plans or differentiates in that subject. It's how the skill knows what "good" looks like in a discipline before it writes a word of lesson. Concretely, per-subject guidance is where things like these belong:

• What counts as rigor in this subject — the kind of thinking a strong lesson should demand, which differs sharply from subject to subject. "Rigor" in math is not "rigor" in social studies.
• How the subject is actually taught — the instructional moves and lesson shapes that fit the discipline, rather than a one-size template stretched over everything.
• What student-facing materials should look like here — the problem set, the close-reading passage, the investigation, the primary-source analysis; each subject's real artifacts, not a generic worksheet.
• How to scaffold within the subject — the supports that help a struggling reader are not the supports that help a struggling problem-solver, and the file captures that difference.
• Which pitfalls to avoid — the subject-specific ways a lesson goes subtly wrong that a generalist wouldn't flag.

Split those out by subject and the skill stops being a generalist doing a passable impression of a subject teacher and starts behaving like it was written by someone who teaches that subject.

Why one-size-fits-all is the wrong size

Consider what a single generic reference would force. To cover four subjects in one file, every instruction has to be vague enough to be true everywhere — which makes it useless anywhere. "Include an engaging activity" is true for math and ELA and science, and tells the skill nothing about what an engaging math activity is versus an engaging ELA one. Genericness and specificity trade off directly; you cannot have guidance that is both broadly applicable and sharply useful in one file.

Splitting by subject dissolves the trade-off. math.md can be as specific as it wants about mathematical practice without worrying about whether it also makes sense for a poetry unit. ela.md can be opinionated about close reading and text-dependent questions without diluting itself to also cover lab investigations. Each file gets to be maximally specific because it only has to be right about one subject.

A generic lesson prompt has to be vague enough to fit every subject, which means it fits none of them well. Per-subject reference files let each one be sharp — and sharpness is what turns a plausible lesson into a good one.

There's a deeper reason this design is sound, and it shows up in how the skills were evaluated. The evaluation framework carries per-subject rubrics — ela, math, science, social_studies, plus a shared set — that mirror the reference files one-to-one. Authoring guidance and evaluation criteria are split along the same subject lines. That symmetry is not an accident: the team held each subject to its own standard on both the writing side and the grading side, which is only coherent if you accept that the subjects are genuinely different crafts.

The fifth file: learning-commons-kg.md

Alongside the four subject files sits learning-commons-kg.md, and it does a different job. Where the subject files carry how to teach a discipline, this file carries what to teach it toward — it's the skill's guide to using the Learning Commons Knowledge Graph connector, which supplies academic standards across all 50 states and the learning progressions beneath each standard.

The pairing is the point. Subject expertise without standards gives you a pedagogically sound lesson pointed at nothing in particular; standards without subject expertise give you a standards-tagged lesson that ignores how the subject is actually taught. Together, the subject file and the KG file let the skill plan a lesson that is both taught the right way and aimed at the right target, scaffolded along the progression the standard is built from. Neither file alone is enough; the reference pack works because it has both.

What subject grounding unlocks with real curricula

Subject-specific reference files also make it possible to lean on subject-specific curricula. Claude for Teachers brings in trusted curricula precisely where subject rigor matters most — OpenSciEd for science and IM v.360 from Illustrative Mathematics for math. A skill grounded in science.md is positioned to use a vetted science curriculum the way a science teacher would; a skill grounded in math.md can build on a real math curriculum rather than generating problems from scratch. The reference file and the curriculum reinforce each other — the file supplies the disciplinary method, the curriculum supplies vetted material, and the standards connector aligns both to the target.

That's the practical payoff of the whole design. When you ask the skill for a science lesson, you're not getting a generic template with science words in it. You're getting a lesson shaped by science-specific guidance, potentially grounded in a trusted science curriculum, aimed at your state's science standards and scaffolded along the progression beneath them. Swap in social studies and every one of those inputs changes to match.

The lesson for anyone building skills

If you take one architectural idea from the open repo, take this one. Domain-specific work needs domain-specific grounding, and the clean way to provide it is separate reference files per domain rather than one overloaded prompt trying to be all things. It keeps each file sharp, it makes the skill's knowledge auditable — you can read exactly what it believes about teaching math — and it lets you extend coverage one subject at a time. The pattern generalizes far beyond K-12; any skill that spans genuinely different domains benefits from the same split.

For a full walkthrough of how those reference files feed a real plan, read the inside look at the lesson-planning skill; for how the matching rubrics judge the output, see how Anthropic evaluated the skills. And for hands-on tooling in the same spirit, the education agent skills and lesson-plan-studio sit in the AI tutoring category, next to more builder insights.

Part of the Claude for Teachers series. Related: Reading the Rubrics · Inside k12-lesson-planning. Browse AI tutoring skills or more builder insights.