Course site: 42-day curriculum LMS at /learn + /admin - #6
Open
BrianJenney wants to merge 12 commits into
Open
Conversation
Turns this repo into the delivery platform for the RAG & AI Agents course, modeled on the parsity_medical_rag LMS: - /learn: Skool-style course home — 6 weeks × day rows, per-day completion toggles, progress bar, continue-where-you-left-off - /learn/day-NN: lesson pages rendered from curriculum/day-NN.md with interactive islands: ```quiz self-checks, ```ai-prompt copyable Claude prompts, ```visual embedded explainers, mermaid diagrams, <details> hint/solution reveals; Descript videos and Typeform assignment links preserved from the original lessons - /admin: invite students by email (Clerk invitations), pending-invite list, per-day progress matrix, revoke/restore access - Auth: Clerk (invite-only email sign-in); progress: Neon Postgres via an isolated Prisma client (prisma/lms) keyed by day slug - curriculum/: 37 day files (+5 rest days) converted from the curriculum branch's module lessons; README week index is the canonical order; AUTHORING.md documents the format - public/visuals/: 7 interactive concept explainers (vector-search, chunking, reranking, hybrid-search, content-validation adapted from the medical-rag course to this course's domain; word-math and agent-router new) - app/libs/chunking.ts: implement getLastWords so main builds (student-todo-exercises keeps the TODO stub); ESLint ignored during builds since main intentionally carries stub files - scripts/check-student-clean.sh guards the student branch from ever receiving LMS/curriculum paths; setup + deploy docs in docs/LMS-SETUP.md Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
- New 'Beyond plain text' section on day-08: honest framing (the course is text-first because that's the common case), PDF extraction reality (text layer vs OCR, table serialization, figures), multimodal embeddings (CLIP-style shared space), and how it maps onto Pinecone (bring-your-own multimodal vectors, metadata.modality) vs bundled- model databases like Weaviate; adds a quiz, an ingestion-planning AI prompt, and a curated search-verified external reading list (Pinecone guides + CLIP notebook, Unstructured, Voyage, Cohere, Weaviate) - New public/visuals/multimodal-rag.html: two-mode explainer — click the elements of a schematic PDF page to see how each becomes a Pinecone record (paragraph, table, figure, scan/OCR), and a shared text+image embedding space where a text query retrieves an image Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
- 5 interview-prep lessons (interview-01..05) converted from the 18-interview-prep module: playbook overview, signature story, tradeoff opinions, RAG system design, live practice — with mock-interview AI prompts, quizzes on the design/tradeoff sessions, and both source videos + the Typeform mock-interview link preserved - Gated per student: Student.interviewUnlockedAt (null = locked, default); /admin progress matrix gets a 🎤 Locked/Unlocked toggle per student - /learn shows the section as a locked teaser card until unlocked, then as a normal lesson list with progress; direct URLs redirect to /learn while locked - Parser: README "## Interview prep" section drives the lesson list (getInterviewLessons/getInterviewLesson in lib/lms/curriculum.ts) - Docs updated (LMS-SETUP: the toggle + schema note). Deploy note: run `yarn lms:push` once after pulling — adds the nullable column, non-destructive Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
- /learn header nav: "❓ Ask a question" (the async-questions Typeform, for concept questions and curriculum errors), "🎤 Interview Prep" (only once unlocked for that student), and an explicit Sign out next to the Clerk avatar (also on /admin) - New ```order interactive island (OrderSteps.tsx): steps presented shuffled, students tap them into sequence, per-position ✓/✗ feedback with retry; deterministic seeded shuffle so SSR/CSR agree; documented in AUTHORING.md - Placed on day-01 (RAG flow), day-08 (ingestion pipeline), day-22 (RAG agent runtime flow — mirrors the five TODO steps students implement right after) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
Admin — student API keys (mirrors the medical-rag cohort tooling, now in-app with server actions): - New "API keys" section in /admin: mint a budget-capped LiteLLM proxy key per student ($10 / 60 days by default, configurable via LITELLM_KEY_BUDGET_USD / LITELLM_KEY_DURATION_DAYS), live spend/budget from /key/info, expiry display, "+$" budget bump (spend preserved), Revoke (kills the key on the proxy), ✉️ Send (prefilled mailto with the key + OPENAI_BASE_URL instructions), copy-key button - lib/lms/litellm.ts: thin server-only layer over the proxy's /key/generate, /key/update, /key/delete, /key/info; section shows a setup note when LITELLM_PROXY_URL/LITELLM_MASTER_KEY are unset - Student model gains apiKey/apiKeyBudget/apiKeyMintedAt/apiKeyExpiresAt (nullable — `yarn lms:push` after pulling) Lessons — two new interactive islands (ephemeral, reset on reload): - ```scenario (Scenario.tsx): workplace "what do you say?" role-play — a manager/PM/dev asks a nebulous question, student picks the reply they'd give, gets a graded verdict (best/ok/weak) with staff-engineer feedback, can reveal how the other replies land; 7 authored across the course: fine-tune-vs-RAG (d12), tool-calling for and against (d31 ×2), stale-docs index updates (d11), few-shot before fine-tune (d16), golden-set evals (d30), voice-vs-knowledge (d20) - ```match (MatchPairs.tsx): tap-to-match pairs with lock-in on correct; chunking-strategy-per-content (d08) and retrieval-mode-per-query (d24) - AUTHORING.md documents both formats with writing rules Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
- curriculum/day-00.md "Start Here — How to Win This Program", listed as Day 0 in the Week 1 index: why the video homework / Feynman Technique is the whole game, the mentor playbook (check-in as metronome, show what you built, ask for push-back, interview them, topics list, and the reach-out path if you don't have a mentor yet), the Slack "2% cheat code" on being present and sharing, momentum-beats-motivation habit guidance, the honest living-curriculum expectations (frequent updates, hiccups happen and how to flag them, text-first on purpose, weekly live sessions + mentors are the product, foundations over tools), a Day 0 checklist, and two AI prompts (success-plan interview, mentor-session rehearsal); day-01's getting-help section now points here instead of duplicating it - New ```blanks island (FillBlanks.tsx): ___N___ markers in code become slots, option pills per blank, per-blank feedback with explanations on misses; placed on day-16 (temperature per job) and day-18 (the selector's zod schemas) - New ```try-it island (TryIt.tsx + /api/lms/try): students run small real API calls with their own class key (from /admin key minting; localStorage only, relayed for exactly one call, models + token caps pinned server-side, Clerk-auth required): embedding-similarity (day-02), temperature 0.0-vs-1.4 (day-16), the selector with a strict JSON schema (day-18), and a live prompt-injection demo with a server-fixed poisoned document (day-34) - AUTHORING.md documents both new islands; validator covers marker/blank parity, answers present in options, and known try-it kinds Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
- Curriculum alignment audit (scripted): all 26 Descript videos from the
source curriculum's mapped lessons are on the correct days with none
invented or dropped; the 2 interview-prep videos are on their pages;
day structure matches DAY-SCHEDULE exactly. Site = source + supplements.
- New ungated bonus-lesson mechanism (slug prefix bonus-, README
"## Bonus lessons" section): "Bonus — Optional Labs" list on /learn,
★ Optional lab page chrome, progress toggles work
- curriculum/bonus-bible-chunking.md: the medical-rag Bible chunking
homework ported to this stack — download KJV from Project Gutenberg,
provided loadVerses() parser inline (no repo/student-branch changes
needed), watch fixed-size slicing fail, strategy tradeoff table +
chunking visual, embed/upsert skeleton matching the upload-text route
(512 dims, PINECONE_INDEX=bible-kjv), verification via searchDocuments
("who is my shepherd?" → Psalm 23 with reference), quiz, 2-3 min video
posted to Slack, defend-my-strategy AI prompt; linked from day-08
- Fix: inline-code chip styling was applying inside dark <pre> blocks,
making all block code in lessons render as unreadable light-on-light
bars; chips now scope to :not(pre) > code only
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
Styling: - Palette sweep across the LMS: indigo → blue-600 primary (now matching the visuals' #2563eb accent), AI-prompt blocks violet → cyan, interview-prep accents violet → sky — no more purple "AI colors" - Cascade fix: the chat app's retro global styles (links, tables, form elements) now exclude the LMS via :not(.lms *) instead of being counter-reset inside .lms — the resets were beating Tailwind utilities and stripping button fills; the LMS is now fully utility-owned - tailwind.config content now includes components/ (classes used only in components were silently missing from the build) Easter eggs (all cosmetic, zero persistence): - Konami code (↑↑↓↓←→←→BA) on /learn → "vector mode": the page becomes a drifting 2-D embedding space of course vocabulary for 12 seconds - Styled console greeting for anyone who opens devtools, with a hint at the cheat code - Confetti burst (site-blue palette) when marking a day done; a bigger one on day-42, whose header also carries a quiet Hitchhiker's nod - Rest-day 🌴 rows sway on hover Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD
… polish Turn the branch into the deployed course site at ai.parsity.io. Auth & access - Clerk sign-in/up routes, middleware gating (/learn + /admin private, landing + auth pages public), production instance wiring - Isolate LMS Prisma into its own schema; env-driven admin allowlist Landing page - Replace the RAG demo at / with a public animated landing (hero, embedding space, RAG pipeline, topics, mentorship/feedback/guest-speaker/interview-prep cards, humans-in-the-loop), reduced-motion safe; new SVG favicon Deploy correctness - Bundle the Prisma query engine into serverless funcs (outputFileTracing + binaryTargets) to fix PrismaClientInitializationError on Vercel - Make ensureStudent idempotent/race-safe (P2002) and email-collision safe - Lazy Pinecone client so builds don't need the key - Bump Next.js to 15.5.x (middleware-bypass CVE) Lessons - Syntax-highlighted code (rehype-highlight), external links open in new tab, mark-done at top and bottom of each lesson Curriculum - Drop decorative emoji, remove unknown time estimates, add class-API-key note to day 5, reframe chunking around retrievability (not token limits), surface the Bible lab as optional-but-encouraged Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add a dedicated "MCP in Production" lesson (three primitives, tool-surface design, auth/permissions, threat modeling, transports) as a going-further day, and make the curriculum freely reorderable. - Day numbers are now COMPUTED from position in the README week index, not hardcoded. Adding, moving, or removing a lesson is a one-line edit; the filenames student progress is keyed on are never renamed. Week day-ranges compute too. Verified the existing numbering is unchanged (day-08 -> Day 8, rests at 7/14/21/28/35, etc.). - New lessons may use semantic slugs instead of day-NN. - The Day 42 easter egg is now slug-based, so it survives renumbering. - Documented the add/move/remove workflow in AUTHORING.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Public "Would You Pass an AI Engineering Interview?" quiz at /ai-interview-quiz: 7 scenario questions drawn from real interviews, server-scored into 3 buckets, email (+ optional phone) capture, Calendly CTA. Leads stored in a new QuizLead table (lms schema); routes public in middleware. - Admin eviction rewired from Clerk ban to delete. banUser is a paid-plan feature and errors on free, so "Remove" now deletes the account, revokes the student's proxy key, and drops their LMS row + progress (behind a confirm). Removes the now-dead ban/restore/"revoked" UI. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Move the chunking corpus to 1536 dimensions, with a "Why 1536 here" section teaching the 512-vs-1536 tradeoff (dimension count is a knob, not a constant). - Add a first-class "Retrieve from your index" step with a 1536-dim query path, turning the dims change into a lesson: the course's 512-dim searchDocuments throws against a 1536 index — query dims must match index dims. - Optional generation stretch (retrieved verses -> cited answer). - Name structure-aware chunking explicitly; add a retrieval reflection to the video. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
BrianJenney
added a commit
that referenced
this pull request
Aug 3, 2026
The Inspector and claude mcp add commands both spawned 'npx tsx', which this project doesn't depend on. When the npx fetch fails the Inspector reports 'Command not found, transports removed' — an error about the runner that reads like a broken server. Documented as failure #6.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Course site: the 42-day RAG & AI Agents curriculum, on the web
Turns this repo into the delivery platform for the course, modeled on the
parsity_medical_ragLMS. Students log in, follow the curriculum day by day, and toggle days complete; admins invite students, mint their API keys, and track everyone's progress.What's in it
Student experience —
/learncurriculum/day-NN.mdwith interactive islands (all ephemeral, only day-completion persists):```quizself-checks (34) ·```ordertap-the-steps-in-order (3) ·```matchtap-to-match pairs (2)```scenarioworkplace role-play (7): a manager/PM/dev asks a nebulous question ("why don't we just fine-tune?", "we should/shouldn't use tool calling", "these docs are stale"), student picks the reply they'd give, gets graded best/ok/weak feedback with staff-engineer reasoning```ai-promptcopy-into-Claude prompt blocks (85) ·```visualembedded interactive explainers (9) ·<details>hint/solution reveals (86) · mermaidstudent-todo-exercisesAdmin —
/adminInfrastructure
curriculum/README.mdweek index = canonical order;AUTHORING.md= full format spec for every island type)Repo housekeeping
getLastWordsimplemented somainbuilds (18/18 chunking tests pass; student branch keeps its stub); ESLint skipped during builds (main carries stub files),tscstill gatesscripts/check-student-clean.shfails if any LMS/curriculum path lands on the student branchVerified empirically:
next buildfails without a Clerk publishable key. Provision first, then merge:NEXT_PUBLIC_CLERK_PUBLISHABLE_KEY,CLERK_SECRET_KEYLMS_DATABASE_URL, thenyarn lms:pushonce (also required after this PR's schema additions — nullable columns, non-destructive)LMS_ADMIN_EMAILS=brian@parsity.io,NEXT_PUBLIC_APP_URL=<prod domain>LITELLM_PROXY_URL,LITELLM_MASTER_KEY(+LITELLM_KEY_BUDGET_USD,LITELLM_KEY_DURATION_DAYS) — the section shows a setup note if unset, nothing breaksFull walkthrough in
docs/LMS-SETUP.md. The chat app at/is untouched.Verification done
next buildpasses with env vars (no-env failure mode documented above);yarn test:chunking18/18/learn//admin🤖 Generated with Claude Code
https://claude.ai/code/session_01SHrJeDcH8rmi1eabHbjEzD