Six steps to mastery, every lesson.
Every ArthurAI™ lesson follows the same six-step structure. The structure isn't arbitrary — it traces from the cognitive-load research and worked-example research into a generative pattern the AI can produce reliably and the educator can review quickly.
- 1
Introduction
Prompt key:
lesson_introductionThe Introduction sets context and orients the learner. It states what the lesson is about, where it sits in the broader curriculum, and what the learner will be able to do at the end. It is short on purpose — front-loading too much content here violates cognitive-load principles. The Introduction adapts to the learner's LCP: more big-picture for global learners, more step-preview for sequential learners.
- 2
Key Concepts
Prompt key:
lesson_key_conceptsThe Key Concepts step names the core ideas the learner needs to engage with the rest of the lesson. Three to five concepts maximum. Each concept gets a one-sentence definition and a one-sentence relevance statement. This step is the ground floor — every later step assumes it.
- 3
Detailed Explanation
Prompt key:
lesson_detailed_explanationThe Detailed Explanation is where the lesson does its real teaching work. Worked examples, derivations, structured exposition, diagrams (sanitized through the Mermaid validator before display), math (validated through the KaTeX renderer). Format weighting reflects the learner's LCP — more visual for visual-leaning learners, more verbal exposition for verbal-leaning learners.
- 4
Practice Questions
Prompt key:
lesson_practice_questionsThe Practice Questions step gives the learner active engagement with the concepts. The questions are tied to the lesson scope and labeled by the competency they test. The AI tutor is intentionally disabled during practice questions so it cannot write the answer for the learner. The educator sees the practice-question performance roll up to faculty analytics.
- 5
Real-World Applications
Prompt key:
lesson_real_worldThe Real-World Applications step grounds the concepts in concrete contexts the learner can recognize. For a younger learner, this might be a familiar everyday situation; for a higher-ed or vocational learner, an industry or research context. The step matters because applied connection is the strongest signal that the learner has actually understood, not just performed.
- 6
Summary
Prompt key:
lesson_summaryThe Summary recaps the key takeaways and review points. It is short, scannable, and structured so the learner can return to it before an assessment. Summaries become the spine of spaced-practice review later in the program.
- 1. Trigger. A learner navigates to a lesson step that has not been generated yet (or an educator triggers regeneration explicitly).
- 2. Prompt assembly. The platform pulls the active prompt from the database (each of the six steps has its own prompt key, versioned, with admin-edited templates). Variables are substituted from context: course name, language, lesson scope, the learner's LCP summary.
- 3. Model invocation. The prompt is routed to the configured model (selection rules document the default and fallbacks per scenario; production routing today centers on Azure OpenAI / OpenAI for tutoring and lesson generation, with provider rotation in roadmap).
- 4. Sanitization. Output is passed through KaTeX math validation and the Mermaid diagram repair pipeline. Schema is validated against the expected step structure. Citations (where applicable) are checked.
- 5. Cache. The generated step is cached so re-visits do not regenerate. The cache is keyed by lesson + step + LCP-summary hash.
- 6. Display. The learner sees the step. The educator can review, edit, or regenerate before publishing the lesson to the cohort. The default workflow is: AI generates, educator attests, learner sees.
Math, diagrams, and content sanitization
- KaTeX math. AI-generated math is validated through KaTeX before display. Malformed expressions trigger a regeneration with an explicit math-formatting clarification.
- Mermaid diagrams. AI-generated diagrams are validated and, where possible, repaired through the Mermaid validator. Non-renderable diagrams are removed from the output rather than shown broken.
- Markdown. Lesson content is rendered through the React Markdown pipeline with GFM, math, and Mermaid extensions. Output is sanitized through DOMPurify before injection.
- Schema enforcement. Each step has an expected JSON shape; schema-mismatched output is regenerated rather than rendered.
See also: the LCP and how it shapes generation · AI tutor architecture · assessments.