Skip to content

Latest commit

 

History

History
940 lines (741 loc) · 30.8 KB

File metadata and controls

940 lines (741 loc) · 30.8 KB

AI PM Technical Fluency - Session Plans

Last updated: February 4, 2026

See MY_PRODUCT.md for how concepts apply to your product.


Teaching Philosophy

For non-dev PMs who want to learn:

  1. Start with WHY — Explain the problem before the solution
  2. Visual first — Diagrams, Console experiments, real examples
  3. Your product — Every concept ties to YOUR product (see MY_PRODUCT.md)
  4. Console over code — Use Anthropic Workbench whenever possible
  5. Simple Python only if needed — 10 lines max, explain each line, YOU run it
  6. Check understanding — Pause and confirm before moving on

Current AI Landscape (as of Jan 2026)

Latest Models

Model Released Best For Price (in/out per MTok)
Claude Opus 4.5 Nov 2025 Most capable, widely used (recommended) $5 / $25
Claude Sonnet 4.5 Sep 2025 Good balance of speed/cost $3 / $15
Claude Haiku 4.5 Oct 2025 Speed, simple tasks, routing $1 / $5

Key New Features (2025-2026)

  • Extended Thinking — Claude shows step-by-step reasoning before answering
  • 1M token context — Sonnet 4.5 supports 1 million tokens (beta)
  • Interleaved thinking + tools — Claude can think between tool calls
  • MCP (Model Context Protocol) — Standard for connecting AI to tools

Phase 1: Foundations (Sessions 1-6)

Session 1: LLM Architecture & API Fundamentals ✓

Concept: How LLMs work, tokens, parameters Exercises:

  • Temperature slider (0 vs 1)
  • System prompts (surfer bro vs formal)
  • Token counting
  • Hallucination demo (no grounding vs grounded) Key Tradeoff: Creativity vs consistency (temperature)

Session 2: Prompt Engineering Patterns ✓

Concept: Few-shot, chain-of-thought, structured output Exercises:

  • Few-shot: show example headlines → get similar style
  • Chain-of-thought: reason through audience/pain/benefit
  • Structured output: get JSON { headline, subheadline, cta }
  • Combine all three Key Tradeoff: Quality vs cost (chain-of-thought uses more tokens)

Session 3: Embeddings & Semantic Similarity ✓

Concept: Text → numbers, meaning positions, similarity Exercises:

  • TensorFlow Embedding Projector (king → queen visualization)
  • Chat API vs Embedding API comparison
  • Conceptual: how headline search would work Key Tradeoff: Generate new vs find existing

Session 4: Vector Databases Deep Dive ✓

Concept: Storing and searching embeddings at scale Before exercises, teach:

  • Record structure: id + values (embedding) + metadata
  • Show example: { "id": "h1", "values": [0.9, 0.8, ...], "metadata": { "text": "Save 50%..." } }
  • Key insight: "The numbers are for searching, the metadata is what you get back"
  • Similarity scores: "Cosine similarity ranges from 0 to 1. Higher = more similar." Exercises:
  • Explore Pinecone console (free tier)
  • Create index, understand configuration
  • Simple Python script to add/search data (explain each line, student runs it) Key Tradeoff: Speed vs accuracy (approximate vs exact search)

Session 5: Document Processing & Chunking

Concept: Breaking documents into pieces for embedding

Start with WHY (no code):

  • Show a full landing page as a wall of text (500+ words)
  • Ask: "Can you embed this whole thing?" → No, embedding models have limits (~8K tokens)
  • Even if you could, searching a giant blob returns... the giant blob. Not helpful.
  • So we need to break it into meaningful chunks

Visual exercise (no code):

  • Take a real example from your product (or mock one up)
  • Print it out / look at it together
  • Ask: "Where would YOU cut this into pieces?"
  • Natural breaks: headline, hero section, features list, testimonials, CTA
  • Discuss: What makes a good chunk?

Console experiment:

  • Paste a SMALL chunk (just headline + subheadline) → ask Claude: "What is this page about?"
  • Paste a LARGE chunk (entire page) → ask same question
  • Paste a MEDIUM chunk (one section) → ask same question
  • Observe: Which answer is most useful? Why?

The tradeoff visual:

SMALL CHUNKS (50 words):
✓ Precise matches
✗ Loses context — "50% off" but off WHAT?

LARGE CHUNKS (500 words):
✓ Full context
✗ Matches too broadly — searching "discount" returns whole page

OVERLAP (chunks share 10-20%):
✓ Preserves context across cuts
✗ Stores more data, costs more

Apply to Your Product: (see MY_PRODUCT.md)

  • A marketer searches "testimonial about fast delivery"
  • Which chunking strategy finds the right section?
  • Discuss: For landing pages, what's the right chunk size?

Key Tradeoff: Chunk size — small = precise but no context, large = context but noisy

Python (only if time, 10 lines max):

  • Simple script that splits text by paragraphs
  • Student runs it themselves
  • Show the output chunks

Session 6: Building Your First RAG Pipeline

Concept: Retrieval Augmented Generation — the full loop

Start with WHY:

  • You've learned: embeddings (Session 3), vector DB (Session 4), chunking (Session 5)
  • Now we connect them: user asks question → find relevant chunks → Claude answers using those chunks
  • This is how you stop hallucination — Claude only uses YOUR data

Visual: The RAG Flow

USER: "Show me landing pages about free trials"
         ↓
    [1. EMBED THE QUERY]
         ↓
    "free trials" → [0.2, 0.8, 0.1, ...]
         ↓
    [2. SEARCH VECTOR DB]
         ↓
    Find chunks with similar embeddings
         ↓
    [3. RETRIEVE TOP MATCHES]
         ↓
    "Start your free trial today..."
    "No credit card required..."
    "Try it free for 14 days..."
         ↓
    [4. INJECT INTO PROMPT]
         ↓
    "Using ONLY this context: {chunks}
     Answer the user's question: {query}"
         ↓
    [5. GENERATE RESPONSE]
         ↓
    Claude writes answer using only retrieved content

Console experiment:

  • First: Ask Claude about your product (no context) → watch it hallucinate or say "I don't know"
  • Then: Paste 3 real landing page chunks as context → ask same question → accurate answer
  • This IS RAG, done manually

Failure modes discussion:

Problem What happens Fix
Bad chunks Retrieved content doesn't answer question Better chunking strategy
Wrong matches Semantic search returns irrelevant content Better embeddings, metadata filters
Too few results Not enough context for good answer Increase top_k
Too many results Claude gets confused by contradictions Reranking (Session 10)

Apply to Your Product: (see MY_PRODUCT.md)

  • Walk through: "Generate a landing page for a summer sale"
  • Where would RAG help? (Find similar past pages, retrieve brand guidelines, get approved CTAs)

Key Tradeoff: Retrieval quality vs generation quality — garbage in, garbage out


Phase 2: Agentic AI (Sessions 7-13)

Session 7: Function Calling & Tool Use

Concept: Claude can call functions — not just generate text

Start with WHY:

  • So far Claude just writes text
  • But what if you want Claude to: check inventory? send an email? look up a customer?
  • You define TOOLS (functions), Claude decides when to use them

Visual: How tools work

YOU DEFINE:
  Tool: "check_inventory"
  Input: product_id
  Output: stock count

USER ASKS:
  "Is the blue widget in stock?"

CLAUDE THINKS:
  "I need to check inventory for blue widget"

CLAUDE CALLS:
  check_inventory(product_id="blue-widget")

SYSTEM RETURNS:
  { "stock": 42 }

CLAUDE RESPONDS:
  "Yes, we have 42 blue widgets in stock."

Console experiment (Anthropic Workbench has tool use):

  • Define a simple tool (character counter, word counter)
  • Ask Claude something that requires the tool
  • Watch Claude decide to use it
  • See the tool call in the response

The autonomy question:

  • Should Claude ALWAYS use the tool? Or DECIDE when to use it?
  • Demo both modes in Console

Apply to Your Product: (see MY_PRODUCT.md)

  • Tool ideas: get_brand_colors, check_compliance, fetch_template, get_approved_images
  • When should Claude auto-use vs ask permission?

Key Tradeoff: Autonomy vs control — let Claude decide vs force tool use

Session 8: ReAct Agents & Extended Thinking

Concept: Agents that reason, act, observe, repeat — plus Claude's built-in thinking

Start with WHY:

  • Simple tool use: one question → one tool → one answer
  • But complex tasks need: think → try something → see result → think again → try something else
  • This is the ReAct pattern (Reason + Act)

Visual: ReAct Loop

USER: "Find me a high-converting landing page template for SaaS"

THOUGHT: I should search for SaaS templates first
ACTION: search_templates(category="SaaS")
OBSERVATION: Found 12 templates

THOUGHT: Now I need to filter by conversion rate
ACTION: filter_by_performance(metric="conversion", min=0.05)
OBSERVATION: 3 templates have >5% conversion

THOUGHT: Let me get details on the top one
ACTION: get_template_details(id="saas-starter-pro")
OBSERVATION: {name, conversion_rate, sections, ...}

ANSWER: "I found 'SaaS Starter Pro' — it has 7.2% conversion rate and includes..."

Extended Thinking (2026 feature):

  • Claude can now SHOW its thinking before answering
  • You set a "thinking budget" (tokens for reasoning)
  • More thinking = better answers, but costs more

Console experiment:

  • Enable Extended Thinking in Workbench (if available)
  • Ask a complex question
  • See Claude's <thinking> block before the answer
  • Compare: same question with/without extended thinking

Apply to Your Product: (see MY_PRODUCT.md)

  • Agent that helps marketers: "Build me a landing page for our Q2 campaign"
  • What tools would it need? What's the ReAct loop?

Key Tradeoff: Thoroughness vs cost/latency — more thinking = better but slower/pricier

Session 9: Memory Systems for Stateful Agents

Concept: How agents remember things

Start with WHY:

  • Chat conversations have memory within a session (Claude remembers what you said)
  • But what about across sessions? What about user preferences?
  • AI products need different types of memory

Visual: Memory Types

SHORT-TERM (conversation):
  "You mentioned you prefer blue earlier..."
  → Just keep chat history in context

LONG-TERM (user profile):
  "This user always uses formal tone"
  → Store in database, inject into system prompt

EPISODIC (past interactions):
  "Last month you created a landing page for Product X"
  → Store summaries, retrieve when relevant

Console experiment:

  • Start a conversation, establish preferences
  • Clear the conversation
  • Start new conversation — preferences are gone
  • Discuss: How would you persist them?

Apply to Your Product: (see MY_PRODUCT.md)

  • Marketer uses AI Page Creation weekly
  • What should the agent remember? Brand voice, past pages, favorite templates?
  • How does this change the experience?

Key Tradeoff: Relevance vs recency — old memories might be outdated

Session 10: Advanced RAG Patterns

Concept: Making retrieval smarter

Start with WHY:

  • Basic RAG: embed query → find similar → done
  • But what if the user's query is vague? Or results aren't quite right?
  • Advanced patterns improve retrieval quality

Pattern 1: Hybrid Search

SEMANTIC: "fast shipping" → finds "quick delivery", "express fulfillment"
KEYWORD: "fast shipping" → finds exact phrase matches

HYBRID: Combine both scores → best of both worlds

Pattern 2: Query Transformation

USER: "pages that convert well"

TRANSFORM TO:
- "high conversion rate landing pages"
- "landing pages with good performance metrics"
- "successful landing page examples"

Search with ALL variations → more comprehensive results

Pattern 3: Reranking

STEP 1: Retrieve top 20 results (fast, approximate)
STEP 2: Use Claude to rerank by relevance (slow, accurate)
STEP 3: Return top 5 reranked results

Console experiment:

  • Give Claude 10 search results, ask it to rank by relevance to a query
  • See how it reorders them
  • This IS reranking

Apply to Your Product: (see MY_PRODUCT.md)

  • Marketer searches "holiday email" — vague query
  • Which pattern helps most?

Key Tradeoff: Complexity vs accuracy improvement — is 5% better retrieval worth 2x complexity?

Session 11: Corrective & Autonomous RAG

Concept: RAG that checks and fixes itself

Start with WHY:

  • Basic RAG retrieves and generates
  • But what if retrieved content is wrong? Or the answer doesn't actually use it?
  • Self-correcting RAG validates its own work

Pattern: Corrective RAG

1. Retrieve chunks
2. Check: "Are these chunks relevant to the question?"
   - If NO → try different search, broader query
3. Generate answer
4. Check: "Does the answer use the retrieved content?"
   - If NO → regenerate with explicit instruction
5. Check: "Is the answer factually consistent with chunks?"
   - If NO → flag for human review

Console experiment:

  • Give Claude chunks + question
  • Ask it to: (a) answer, then (b) grade whether its answer used the chunks faithfully
  • See self-evaluation in action

Apply to Your Product: (see MY_PRODUCT.md)

  • AI generates a landing page
  • How would you check: Did it follow brand guidelines? Did it use approved copy only?

Key Tradeoff: Quality vs latency — more checks = better quality but slower

Session 12: When RAG Fails — Agentic Search Alternatives

Concept: Real-world cases where RAG underperforms, and how agentic search wins

Start with WHY:

  • You've learned RAG is powerful — but it's not always the answer
  • Anthropic abandoned RAG for Claude Code because it underperformed
  • Sometimes simpler approaches work better

RAG's Hidden Costs:

STALENESS:
  Code changes → Index is outdated → Wrong results
  How often do you re-index? Real-time? Hourly? Daily?

RELIABILITY:
  Embedding quality varies
  "Semantic similarity" doesn't always mean "relevant"

SECURITY:
  Indexed data = stored data = attack surface
  Anthropic deemed their own codebase too sensitive to index

Visual: RAG vs Agentic Search

RAG APPROACH:
  1. Pre-index all content → Vector DB
  2. User query → Embed → Search index → Return chunks

  Problems: Stale, requires infrastructure, security risk

AGENTIC SEARCH APPROACH:
  1. User query → Agent thinks: "What do I need?"
  2. Agent uses tools: grep, glob, file read
  3. Agent iterates: search → read → search more → synthesize

  Benefits: Always fresh, simple tools, no index to secure

Case Study: Claude Code

  • Early versions used standard RAG for codebase search
  • Boris Cherny (Anthropic engineer) shared why they abandoned it:
    • Staleness: Code changes faster than indexes update
    • Reliability: Retrieval quality was inconsistent
    • Security: Even internal indexing was a liability
    • "Vibes": It just didn't feel as helpful
  • Agentic search (grep + glob + iterate) outperformed RAG "by a lot"
  • Result: 70% per-engineer productivity growth

The "Vibes" Metric:

  • Benchmarks can lie — a system can score well but feel wrong
  • Real test: Does it help you accomplish your goal?
  • Sometimes simpler = better user experience

Console experiment:

  • Scenario: Find all mentions of "pricing" in a codebase
  • RAG approach: Embed "pricing" → search → get chunks
  • Agentic approach: grep "pricing" → read files → refine search
  • Compare: Which gives you better, more actionable results?

Decision Heuristic: When to Index vs Search Dynamically

USE RAG WHEN:
  - Massive corpus (millions of documents)
  - Sub-second latency required
  - Content is relatively static
  - Security/privacy isn't critical

USE AGENTIC SEARCH WHEN:
  - Content changes frequently
  - Security is paramount
  - "Good enough" retrieval is fine
  - You want simplicity over infrastructure

Apply to Your Product: (see MY_PRODUCT.md)

  • Template library: RAG might make sense (static, needs speed)
  • Brand guidelines doc: Agentic search? (changes, security matters)
  • Customer's uploaded assets: Which approach?

Key Tradeoff: Infrastructure complexity vs model capability — as models improve, simpler often wins


Session 13: Voice & Multimodal Agents

Concept: Beyond text — images, audio, video

Start with WHY:

  • Claude can SEE images (vision)
  • This unlocks: analyze screenshots, review designs, understand charts
  • For landing pages: "Look at this page and tell me what's wrong"

Console experiment:

  • Upload a landing page screenshot to Claude
  • Ask: "What's the headline? What's the CTA? What could be improved?"
  • Claude describes what it SEES

Apply to Your Product: (see MY_PRODUCT.md)

  • Marketer uploads competitor's landing page: "Make me something like this"
  • Claude analyzes: layout, colors, sections, copy patterns
  • Generates similar output using your product's components

Voice concepts (discussion):

  • Voice → text → Claude → text → voice
  • Real-time voice agents
  • When does voice make sense for your product?

Key Tradeoff: Modality richness vs complexity — vision is powerful but harder to debug


Phase 3: Multi-Agent Systems (Sessions 14-19)

Session 14: Multi-Agent Architecture Patterns

Concept: When one agent isn't enough

Start with WHY:

  • Single agent handles simple tasks
  • Complex tasks need specialists: one researches, one writes, one reviews
  • Like a team, not a solo worker

Visual: Single vs Multi-Agent

SINGLE AGENT:
User → Agent → Response
(Agent does everything)

MULTI-AGENT:
User → Orchestrator → [Research Agent] → [Writing Agent] → [Review Agent] → Response
(Specialists collaborate)

Patterns:

Pattern How it works Best for
Sequential A → B → C Pipelines (research → write → edit)
Parallel A, B, C simultaneously Independent subtasks
Hierarchical Boss agent delegates to workers Complex projects

Apply to Your Product: (see MY_PRODUCT.md)

  • "Build a landing page for summer sale"
  • Agent 1: Research past summer campaigns
  • Agent 2: Generate copy options
  • Agent 3: Assemble layout
  • Agent 4: Review for compliance
  • Is this overkill? When is multi-agent worth it?

Key Tradeoff: Specialization vs coordination overhead

Session 15: CrewAI for Role-Based Teams

Concept: Agents with personas — roles, goals, backstories

Start with WHY:

  • Generic agents are good
  • But agents with ROLES perform better: "You are a senior copywriter..."
  • CrewAI makes this pattern easy

Visual: CrewAI Crew

CREW: Landing Page Team

AGENT: Research Analyst
  Role: Find relevant data and examples
  Goal: Provide comprehensive research
  Backstory: "10 years in market research..."

AGENT: Copywriter
  Role: Write compelling copy
  Goal: Maximize conversions
  Backstory: "Award-winning direct response writer..."

AGENT: Editor
  Role: Review and polish
  Goal: Ensure quality and brand consistency
  Backstory: "Former brand manager at Fortune 500..."

TASK FLOW: Analyst → Copywriter → Editor → Final output

Conceptual exercise:

  • Define roles for your product's AI crew
  • What's each agent's goal? Backstory?
  • How do they hand off work?

Key Tradeoff: Role clarity vs flexibility — strict roles can limit creativity

Session 16: LangGraph for Complex Workflows

Concept: Workflows as graphs — with branches and loops

Start with WHY:

  • Linear workflows: A → B → C
  • But real work has: "If X then do Y, else do Z", loops, human approval steps
  • Graphs let you model this complexity

Visual: Graph Workflow

        [Start]
           ↓
      [Generate Draft]
           ↓
      [Review Quality]
         /    \
    [PASS]    [FAIL]
       ↓         ↓
   [Publish]  [Revise] ──→ [Review Quality] (loop)

Key concepts:

  • Nodes: Steps in the workflow (generate, review, publish)
  • Edges: Connections between steps
  • Conditional edges: If/then branching
  • Human-in-the-loop: Pause for approval

Apply to Your Product: (see MY_PRODUCT.md)

  • Landing page creation workflow
  • Where do you need human approval?
  • Where can you loop (revise until approved)?

Key Tradeoff: Expressiveness vs debuggability — complex graphs are hard to troubleshoot

Session 17: OpenAI Agents SDK Deep Dive

Concept: How OpenAI approaches agents — compare to Anthropic

Start with WHY:

  • Different vendors, different philosophies
  • Understanding both helps you choose — and avoid lock-in
  • OpenAI SDK vs direct Claude API — tradeoffs

Compare:

Aspect OpenAI Agents SDK Claude API + tools
Abstraction High (SDK handles orchestration) Low (you build it)
Flexibility Constrained to their patterns Build anything
Lock-in High (their SDK, their models) Lower (swap models easier)

Discussion:

  • When does an SDK help vs constrain you?
  • For your product: build custom or use a framework?

Key Tradeoff: Ecosystem lock-in vs specialized features

Session 18: Google ADK for Enterprise Agents

Concept: Enterprise requirements — compliance, scale, integration

Start with WHY:

  • Startups can move fast and break things
  • Enterprises need: audit logs, data residency, SSO, compliance
  • Google ADK designed for this

Enterprise requirements checklist:

  • Data stays in region (GDPR, etc.)
  • Audit logs for all AI decisions
  • Role-based access control
  • Integration with existing identity systems
  • SLAs and support

Discussion:

  • Which of these matter for your enterprise clients?
  • How do you evaluate vendors on enterprise readiness?

Key Tradeoff: Enterprise features vs agility — compliance slows things down

Session 19: MCP & Standardized Tool Calling

Concept: Model Context Protocol — universal standard for AI tools

Start with WHY:

  • Every AI system invents its own tool format
  • MCP: one standard for connecting AI to tools
  • Like USB for AI — plug and play

Visual: MCP Architecture

[Claude Desktop] ←→ [MCP Server: File System]
                 ←→ [MCP Server: Database]
                 ←→ [MCP Server: Your API]

One protocol, many connections

Demo (if Claude Desktop available):

  • Show MCP servers in Claude Desktop
  • Connect to a file system MCP
  • Watch Claude read/write files

Apply to Your Product: (see MY_PRODUCT.md)

  • MCP server that exposes: templates, brand assets, compliance rules
  • Any AI tool (Claude, GPT, etc.) can connect
  • Build once, use everywhere

Key Tradeoff: Standardization vs innovation speed — standards slow early adoption


Phase 4: Production (Sessions 20-25)

Session 20: Evaluation Frameworks That Matter

Concept: How to know if your AI actually works

Start with WHY:

  • You shipped an AI feature — is it good?
  • "Feels good" isn't enough — you need metrics
  • But AI is hard to evaluate — outputs vary

Evaluation types:

Type How it works Pros Cons
Human eval People rate outputs Gold standard Expensive, slow
Automated Model grades model Fast, cheap Can miss nuances
A/B testing Compare versions Real user signal Needs traffic

Key metrics for AI features:

  • Task completion rate (did it work?)
  • User satisfaction (did they like it?)
  • Accuracy (was it correct?)
  • Latency (was it fast enough?)

Console experiment:

  • Generate 5 landing page headlines
  • Grade them yourself (1-5)
  • Ask Claude to grade them (1-5)
  • Compare: Do you agree?

Apply to Your Product: (see MY_PRODUCT.md)

  • How do you measure if AI Page Creation is "good"?
  • What's the bar for shipping?

Key Tradeoff: Evaluation cost vs confidence — more testing = more confidence but slower shipping

Session 21: Safety, Guardrails & Responsible AI

Concept: Preventing AI from doing harm

Start with WHY:

  • AI can generate bad stuff: false claims, offensive content, legal risks
  • Your job as PM: define guardrails before shipping
  • Balance: too strict = useless, too loose = dangerous

Guardrail types:

INPUT GUARDRAILS:
  Block: "Write me malware"
  Allow: "Write me a landing page"

OUTPUT GUARDRAILS:
  Block: Medical claims, price guarantees, fake testimonials
  Allow: Marketing copy, product descriptions

TOPIC GUARDRAILS:
  Block: Politics, competitors, legal advice
  Allow: Your product, general marketing

Console experiment:

  • Try to get Claude to write something inappropriate for a landing page
  • See how Claude refuses
  • Discuss: What guardrails does your product need?

Apply to Your Product: (see MY_PRODUCT.md)

  • AI can't promise "100% guaranteed results"
  • AI can't use competitor names
  • AI must follow FTC guidelines for testimonials
  • How do you enforce these?

Key Tradeoff: Safety vs utility — over-filtering kills usefulness

Session 22: RLHF & Alignment Deep Dive

Concept: How Claude became helpful and safe

Start with WHY:

  • Raw language models are chaotic — just predict next word
  • RLHF (Reinforcement Learning from Human Feedback) trains them to be helpful
  • Understanding this helps you understand Claude's behavior

Visual: RLHF Process

1. PRE-TRAINING: Read the internet, learn language
2. FINE-TUNING: Train on helpful examples
3. RLHF:
   - Generate responses
   - Humans rank: "This response is better than that one"
   - Model learns from rankings
4. CONSTITUTIONAL AI (Anthropic):
   - Model self-corrects based on principles
   - "Would this response be harmful?"

Discussion:

  • Why does Claude refuse some requests?
  • Why does Claude ask clarifying questions?
  • How does this affect your product design?

Key Tradeoff: Helpfulness vs harmlessness — can't maximize both

Session 23: System Design for AI Products

Concept: Architecture that works at scale

Start with WHY:

  • Prototype works on your laptop
  • Production needs: multiple users, reliability, cost control
  • Architecture decisions now affect you later

Visual: AI Product Architecture

USER → [Your App] → [API Gateway] → [Claude API]
                         ↓
                  [Vector DB (Pinecone)]
                         ↓
                  [Your Database]
                         ↓
                  [Caching Layer]

Key decisions:

Decision Options Tradeoff
Sync vs async Wait for response vs background job UX vs complexity
Caching Cache responses vs always fresh Cost vs accuracy
Fallbacks What happens when Claude is down? Reliability vs cost

Draw together:

  • Your product's AI feature architecture
  • Where's the vector DB? Where's Claude? Where's caching?
  • What's the critical path?

Key Tradeoff: Latency vs accuracy vs cost — pick two

Session 24: Fine-tuning vs RAG vs Agentic Search Decision Framework

Concept: When to fine-tune, when to RAG, when to search dynamically, when to prompt engineer

Start with WHY:

  • You now know FOUR approaches to giving Claude knowledge:
    1. Prompt engineering: Just tell Claude what you want
    2. Agentic search: Let Claude dynamically search with tools
    3. RAG: Pre-index content, retrieve at query time
    4. Fine-tuning: Train Claude on your data permanently
  • Each has different strengths — choosing wrong = wasted effort

The Four Options Spectrum:

SIMPLEST ←————————————————————————————————→ MOST COMPLEX

Prompt Engineering → Agentic Search → RAG → Fine-tuning
     ↓                    ↓            ↓         ↓
  "Just ask"         "Search live"  "Index it"  "Train it"

Decision Framework:

START HERE: Can you just prompt it?
  YES → Use prompt engineering (few-shot, system prompts)
  NO → Continue...

Is the content SENSITIVE or SECURITY-CRITICAL?
  YES → Agentic search (no index to secure)
  NO → Continue...

How DYNAMIC is the content?
  Changes hourly/daily → Agentic search
  Changes weekly/monthly → RAG with re-indexing
  Rarely changes → RAG or fine-tuning

How LARGE is the corpus?
  Small (<100 docs) → Agentic search
  Medium (100-10K docs) → RAG
  Massive (>10K docs) → RAG with good chunking

Latency requirements?
  Sub-second required → RAG (pre-indexed)
  2-5 seconds OK → Agentic search

Is it STYLE or KNOWLEDGE?
  Style (write like us) → Fine-tuning or few-shot
  Knowledge (facts about products) → RAG or agentic

When Agentic Search Beats RAG:

Scenario Why Agentic Wins
Codebase search Code changes constantly, security matters
Internal docs Freshness > speed
Exploratory queries Need to iterate and refine
Small corpus Index overhead not worth it

When RAG Still Wins:

Scenario Why RAG Wins
Customer support KB Massive, needs instant answers
Product catalog Structured, changes predictably
Legal/compliance docs Needs audit trail of what was retrieved

Hybrid Strategies:

STABLE + DYNAMIC DATA:
  RAG for: Product specs, pricing (changes weekly)
  Agentic for: Inventory, real-time data

TIERED RETRIEVAL:
  1. Fast RAG search (80% of queries)
  2. Fall back to agentic search (complex queries)

Apply to Your Product: (see MY_PRODUCT.md)

  • Brand guidelines: Prompt engineering + few-shot? (small, stable)
  • Template library: RAG (large, structured, needs speed)
  • Customer's uploaded assets: Agentic? (varies per customer, security)
  • Compliance rules: Prompt + RAG hybrid?

Cost/Benefit Analysis Exercise:

Approach Setup Cost Maintenance Query Cost Latency
Prompt Low Low Medium Fast
Agentic Low Low High (more tokens) Medium
RAG High Medium Low Fast
Fine-tune Very High High Low Fast

Key Tradeoff: Customization vs maintenance vs simplicity — there's no universal best answer

Session 25: Capstone Integration

Concept: Put it all together — design the full feature

You've learned:

  • Phase 1: How AI works (LLMs, embeddings, vector DBs, RAG)
  • Phase 2: Agents (tools, memory, self-correction, agentic search, multimodal)
  • Phase 3: Multi-agent (teams, workflows, standards)
  • Phase 4: Production (eval, safety, scale, decisions)

Capstone exercise:

  • Design the full AI feature for your product (see MY_PRODUCT.md)
  • Draw the architecture
  • List the components (RAG? Agents? What tools?)
  • Define the evaluation metrics
  • Identify build vs buy decisions
  • Write a 1-page spec

Deliverable: Product spec for AI Page Creation

Quiz: Comprehensive review of all concepts


Session Checklist Template

Each session should include:

  • Start with WHY (no code, explain the problem)
  • Visual or analogy to anchor the concept
  • Hands-on exercise (Console preferred, simple Python if needed)
  • Apply to your product (see MY_PRODUCT.md)
  • Discuss key tradeoff
  • Check understanding before moving on
  • 8-question quiz
  • Update progress