The rubric is the source of truth. Each question stores what counts as correct, what counts as partial, and what counts as wrong, in structured form. The system prompt instructs the bot to score against the rubric only and to explain in the rubric's own language. Without a clear rubric the bot defers to a 'I cannot score this confidently' fallback rather than guessing. Tighten the rubric, the scoring tightens with it.

Yes for many cases. Code that has a deterministic output can be scored by an actual code-execution tool function (a sandbox runs the learner's submission and compares the output to expected). For conceptual code questions ('what does this return') the bot reads the snippet, applies language rules, and matches against the rubric. For long-form code (write a function that does X), use the rubric to specify required behaviours, edge cases, and antipatterns, and the bot scores against those individually.

It handles image-based questions through multimodal models (GPT-4o, Claude Sonnet, Gemini). The learner can upload a diagram or photo and the bot evaluates against the rubric (label this diagram, identify the misalignment, what is wrong with this UI). For audio and video questions, transcript-first works for most language and music-theory cases. Pure spatial reasoning is still hard for any model, so use sparingly.

Each question in the bank carries a difficulty rating (1-5) and a topic tag. The running score and per-topic accuracy track in wp_sleek_ai_quiz_progress. After each question the bot picks the next from the bank based on a target accuracy band (around 60-70% is the productive-struggle zone for most learners). Three in a row correct triggers a step up, two in a row wrong triggers a step down. Topic gaps surface harder questions on that topic specifically.

Yes, with caveats. The bot can generate questions from a corpus of source material (course pages, a textbook chapter, a video transcript) if you want unlimited practice. Generated questions are inherently less reliable than human-authored ones, so a review queue lets a human approve them before they enter the bank. Most teams use a mix: human-authored core bank for assessed quizzes, model-generated unlimited practice for casual revision.

Per logged-in WordPress user, in a custom table or post-meta against the user record. Each row stores question ID, answer given, score, timestamp, and topic tag. From there you can compute streaks, topic accuracy, and time-to-mastery. For anonymous learners, progress lives in a browser cookie or session storage, with an offer to save by signing up after the first quiz completion. Logged-in is the better path because progress survives device changes.

Yes via SleekAI multibot. Each subject area or course can have its own quiz bot scoped to its URL pattern, with its own question bank and rubric, sharing the underlying progress table. A learner taking the JavaScript quiz, then the Python quiz, sees progress for each topic separately. Leaderboards and certificates can roll up across all quizzes if you want a cross-course recognition system.

It depends on the LMS you compare against. For formal credit-bearing assessment with proctoring, AICC/SCORM compliance, and audit-grade gradebooks, dedicated LMSes (Moodle, Canvas) still win. For informal practice, certification prep, marketing-funnel quizzes, and community-building quizzes, SleekAI's chatbot quiz is generally superior because the explanation quality is so much higher than multiple-choice and the data lives in your own WordPress. Pick the right tool for the assessment stake.