Core concepts

Learn the fundamental concepts of Agent Skills including skills, evaluator rubrics, and how they connect to Knowledge Graph.

How Agent Skills are designed

Concept	Definition
Skill	A folder containing a SKILL.md file with metadata and step-by-step instructions that tell an agent how to complete a specific K-12 teaching workflow. Example: The K-12 Lesson Planning skill routes by subject, grounds in standards, and renders classroom-ready artifacts from a master JSON file.
Reference files	Subject-specific markdown files bundled with each skill that carry pedagogy, curriculum branching, grade-band structures, and output mappings. Example: A math lesson loads references/math.md before drafting; skipping this step is treated as a critical failure.
Bundled scripts	Helper scripts included with each skill to render editable .docx Word documents from structured JSON output. Example: Both lesson skills render teacher plans and student materials from one master JSON via shared render scripts.
Knowledge Graph MCP Server	Optional MCP tools that lets Agent Skills resolve standards, learning components, progressions, and curriculum context at runtime. Example: When connected, the skill calls find_standard_statement before drafting; when not connected, it falls back to general knowledge with a disclosure footer.

How quality is evaluated

Concept	Definition
Rubrics	A CSV of scoreable criteria used to benchmark the quality of skill outputs. Each criterion includes a pass condition, optional conditionality, and design notes. Example: Lesson planning rubrics include shared.csv plus subject-specific files such as math.csv and ela.csv.
P/R/O/M framework	All criteria fall into one of four buckets. The buckets reflect two paired goals: quality (does the output reflect strong pedagogy and appropriate rigor?) and usability (is the output formatted and scaffolded in a way that a real teacher can actually use it?) P — Pedagogy — Pedagogy criteria evaluate whether the output reflects sound instructional design: standards alignment, prerequisite and forward connections, appropriate instructional model, discourse structures, attention to student struggle, and visual/representational choices. Key pedagogical commitments: Curriculum coherence. When a teacher uses a high-quality instructional material (HQIM), outputs should be coherent with that curriculum’s design logic. Discipline-specific instructional models. Math rubrics measure full standards coverage and student discourse. ELA rubrics are grounded in the science of reading. Science rubrics require inquiry, writing, and quantitative reasoning. Social studies rubrics focus on sourcing and argumentation. Anticipating struggle. P criteria require attention to points of difficulty — patterns, why they arise, and teacher moves for each. R — Rigor — Rigor criteria evaluate whether the output maintains grade-level cognitive demand: Grade-level access for all students. No materials simplify below the standard’s cognitive demand without a scaffold that preserves the core challenge. Critical thinking demanded. At least one task requires students to explain reasoning, evaluate a strategy, or connect representations. Student agency. Student-facing materials include at least one open-ended or reflective prompt. O — Output/Formatting — Output criteria evaluate the artifact itself: correct file structure, appropriate length, teacher- vs. student-facing separation, universal design features, and teacher rationale notes. One criterion worth calling out: designed for teacher adaptation. Teacher-facing outputs should note which elements are non-negotiable and why. M — Model Scaffolding — Model Scaffolding criteria evaluate the model’s conversational behavior: whether it asks for missing information before generating, whether it proactively produces student-facing materials, and whether it offers meaningful follow-up options.
LLM as judge	An evaluation method where an LLM scores generated materials against rubric criteria as pass/fail, with explanations. Learning Commons publishes rubrics openly so developers can reuse the same quality bar in their own evaluator harnesses.
Conditional criterion	A rubric criterion that applies only when a specified condition is met (for example, K-5 or ELA-Gr8+). If the condition isn’t met, the criterion is skipped — not failed.

How Agent Skills relate to other tools

Tool	Relationship
Knowledge Graph	Supplies standards, learning components, progressions, and curriculum data that skills draw on through MCP tools. Strongly recommended for accurate outputs.
Evaluators	Measure AI output quality using rubrics such as literacy complexity and feedback quality (includes evaluators, rubrics, and related tooling). Agent Skills evaluator rubrics focus specifically on K-12 instructional artifact quality.

Related topics

Skills

Explore available Agent Skills and their published eval rubrics.

Evaluating outputs

Benchmark skill outputs against rubrics with LLM-as-judge or human review.