BUILD A GENERAL-PURPOSE LEARNING SYSTEM. Do not blindly impl...

Prompt

BUILD A GENERAL-PURPOSE LEARNING SYSTEM. Do not blindly implement this spec: act like sculptor/poet. Preserve thesis, remove redundancy/features that do not serve it, merge primitives, redesign architecture, add missing mechanisms where clearly useful. NORTHSTAR: learner arrives bored/uncertain/curious → within minutes engaged → flow → willing to sustain effort → able to learn new knowledge/skills flexibly and act. Optimize correct mental-model formation, exact failure detection, useful friction, retrieval, independent capability—not consumption/watch-time/completion. Hook→curiosity→wonder→struggle→aha→retrieval→capability→creation. THESIS Not content library; mental-model construction system. Gym, not library: tight feedback loop around mistakes. Every concept exists because prior problem/constraint demanded it. Lesson=reasoning process, not container. CORE RULES * no abstraction before need: prior situation/context→failure/constraint→necessary concept * working memory: ≤3 new primitives at once * generation>recognition: derive/implement/predict/contrast/reconstruct * scheduled active retrieval, not rereading * evidence>self-report; mechanism>decoration; independent problem solving>completion * personalization changes path/order/depth, never removes capability LEARNER MODEL / ENTRY On first use, probe with few questions/tasks to infer goal, background, existing knowledge, misconceptions, identity/role, preferred entry, cognitive state. Start at actual level; progressively climb to target. Crucial: bridge across domains using learner's existing vocabulary/interests. Example arts→math: do not force artificial subject silo; construct a meaningful bridge so learner can walk into unfamiliar field with joy/flow. 2D personalization = WHO learner is × HOW learner enters graph: bottom-up fundamentals-first vs top-down projects-first. Personas: researcher/engineer/investor/enthusiast/unfiltered; same knowledge, different sequencing/depth. State modes: calm/curious/focus/energized/wonder/overwhelmed/momentum/bored/deep. State changes content+environment+interaction. Bored/~20m→game/video/post; moderate→lesson; deep→lesson+deeper. Avoid dense reading when unwilling; vary routing to avoid staleness. UX / SHELL Grouped nav Study/Papers/Frontier/Media/Plan; low chrome, content protagonist; persistent light/dark; homepage answers where-left-off/what-next/progress/what-relevant and feels like product, not marketing. Glossy/glass, restrained chrome, color as environmental signal, precise diagrams, animation only for mechanisms; UI must never become object of attention. Incremental loading as corpus grows. KNOWLEDGE ARCHITECTURE Course→Chapter→Lesson with filters/progress/deep links/persona ordering/math attachments. Information architecture exposes dependency structure, not just menus. Concept graph = prerequisite DAG + dependency chains + later uses + review state + centrality; always answer WHY learn this / WHERE used. Research graph: foundations→papers→history→frontier; timeline Problem→Invention→Why mattered→What enabled; typed claims/dates; trajectory, not model-name list. LESSON ENGINE Question(problem)→Context(prior state)→Pressure(failure)→Primitives(≤3)→Derivation(constraint forces idea)→Figure(mechanism)→Implementation(operate)→Prediction(commit before result)→Contrast(plausible non-strawman wrong approach)→Retrieval(reconstruct)→Generative Exercises→Check Gate(prove before advance)→Chapter Math(compression)→Reference(prereqs/misconceptions/sources)→Completion. Core: lesson=reasoning process. Math is compression: dedup equations, backlinks, notation tables; learner eventually compresses chapter theory to compact math. Context glossary: popover definition/why/prereqs/formula/figure/lessons, smart position/survives scroll. Goal: basic comprehension not outsourced to LLM chat. Chapter synthesis = first-class lesson: reconstruct concepts, relations, equations, dependency chain; chapter-level compression. FAILURE / MEMORY Mistake ledger = persistent pointer to lesson/exercise/concept/type/frequency/history, not duplicated content. Personalization unit=WHERE learner broke, not WHAT studied. Retrieval mix: unresolved mistakes + due recalls + recent consolidation + re-asked generative tasks + synthesis. Forgot/Hard/OK/Good/Easy→future intervals. Retrieval=active reconstruction. Concept-gap urgency ≈ mistakes×recency/repetitions; scheduler input; optional hard-mode gate. Reminder = pre-generated retrieval task, not “study today.” Memory targets what is forgotten, not what was consumed. INTERACTION / LAB / PLAY Interactive lab executes mechanisms (BPE, softmax/temperature, entropy, Laplace smoothing, sampling, n-gram etc); lesson can push code directly into lab. Understanding mechanism=operate it. Generative playgrounds: real-time simulations; tweak variables, discover formula/mechanism; avoid binary wrong/full-solution dump. Games: timed word-by-word reveal+MCQ, timeout=fail, same mistake ledger; low-friction bored mode; tests meaning construction under pressure, not decorative gamification. PROJECTS / EXECUTION Projects split micro-lessons: goal→concept→starter→reference implementation→checks→commit. Loop learn→implement→verify→commit→evidence. Artifacts matter, not completion records. Hosted execution future: server-side code→phone/tablet; auto-grading→mistake system; write→execute→test→fail→diagnose→retry; isolation/timeouts/quotas/cost limits. Expanded grading: passes tests/fails cases; code gives objective evidence/proof. PLANNING Input horizon/minutes/start date→daily lessons/projects/papers/frontier/review/revision; calendar/.ics export. If workload doesn't fit, explicitly report it; surplus→revision, never fake busywork. RESEARCH / PAPERS / FRONTIER Paper dissection: problem/prior work/idea/architecture/objective/experiment/result/limitation/falsifier/aftermath + predict→reveal; reconstruct researcher reasoning, not summary. Frontier: post-training/architectures/agents/eval/economics/open weights/lab practice; changed items + attributed claims + open questions + sources. Must be continuously updated, not treat current models as endpoint. Researcher Mode: learner attacks open problem first, then sees historical/frontier solution. Interactive paper figures: attention/Chinchilla/Mixtral/MLA etc become manipulable; figure may be argument. MEDIA Curated videos/channels/live filtered posts/user sources. State-sensitive alternative to textbooks but separate from core learning system to prevent passive-content feed. VISUALS / SOURCES / EPISTEMICS Visual pipeline: identify need→retrieve→verify relevance/license→attribute→human review. Generated visuals only conceptual illustration, never fabricated data/charts. Visual correctness>visual usefulness; static figure default. Living knowledge pipeline: auto discover/extract/embed/classify/verify/distill RSS/YouTube/arXiv etc→timelines/term evolution/superseded-by. Disagreement engine: se

A system prompt was added to support web rendering