Foundations
Claude Certified Architect - Foundations
The whole certification as bullets and pictures, with the official exam keywords and the real situations behind each one.
Module 1
AI Fluency: Framework & Foundations
The 4D Framework
- AI Fluency: effective, efficient, ethical, safe
- Four competencies: Delegation, Description, Discernment, Diligence
- Apply across automation, augmentation, agency
- Engagement skills outlast any tool
Important to learn
Why These Four
A cook masters heat and timing, not recipes; 4Ds outlast any stove
Delegation
- Delegation distributes tasks between self, AI
- Generative AI uses probability, not retrieval
- Reliable on common, weak on rare topics
- Needs domain expertise plus AI understanding
Important to learn · exam keywords
unclearTask tool
Splitting chores, keep your judgment, hand out the dependable parts
Description & Discernment
- Description communicates intent to AI
- Discernment evaluates whether output fits
- Prompt engineering blends human, AI principles
- Describe, judge gap, refine: feedback loop
Important to learn
The Description-Discernment Loop
Directing a driver, speak, watch the turn, correct next instruction
Diligence & Conclusion
- Diligence owns ethics, safety, responsibility
- Contexts differ on disclosure, verification
- Four competencies: connected practice, not checklist
- 4D vocabulary spans Developer, Associate, Architect
Important to learn
Context Sets the Expectation
Driver stays responsible even when cruise control steers
Module 2
Building with the Claude API
API Access & Requests
- Route calls through your server, hide API key
- Resend full message history every request
- Temperature adds randomness, never guarantees output
- System parameter shapes role, rejects system=None
Important to learn · exam keywords
Write-p / --print.claude/rules//memory@import
Forgetful storyteller needs the whole tale retold each visit
Prompt Evaluation
- Eval measures effectiveness with dataset scores
- Pipeline iterates on objective numbers
- Model grading needs reasoning, avoids middling 6
- Code grading parses syntax, returns 10 or 0
Important to learn · exam keywords
Writeother + detailstructured error context
A rubric judge scoring many dishes beats one lucky bite
Prompt Engineering
- Iterate: change one thing, measure
- Clear and direct: lead with action verb
- Specific output guidelines doubled the score
- Multishot examples in XML, explain why ideal
Important to learn · exam keywords
custom_idexplicit context passinginput formatsother + detailstructured error context
Turn one radio knob to learn which cleared the static
Tool Use
- Schema, Claude requests, you run, feed back
- tool_result must match tool_use_id
- Loop while stop_reason equals tool_use
- Fine-grained disables JSON validation, handle invalid
Important to learn · exam keywords
tool_useEdit-p / --print@importmultiple matches
Numbered tickets match numbered plates even when jumbled
Retrieval-Augmented Gen
- Chunk ahead, retrieve relevant at query time
- Embeddings and cosine similarity find meaning
- BM25 lexical weights rare terms, exact matches
- Retriever fuses vector and BM25 via RRF
Important to learn · exam keywords
example queriesmultiple matches
A well-read librarian plus a strict exact-match index
Claude Features
- Extended thinking judiciously, only when evals demand
- Image prompting: methodology beats naive count
- Prompt caching needs byte-identical ephemeral breakpoints
- Code execution container network-less, Files API bridges
Important to learn · exam keywords
.claude/rules/Task toolstructured metadataverify_fact
A sealed workshop fed through one materials hatch
Model Context Protocol
- MCP shifts tool burden to servers
- Client bridges transport-agnostic message passing
- mcp.tool decorator auto-generates schema from hints
- Resources expose data, tools perform actions
Important to learn · exam keywords
Readbusiness errorcustom_iddirect executionexample queries
Plug into pre-built sockets the utility already maintains
Agents & Workflows
- Workflows predetermined, agents plan themselves
- Prefer workflows, agents only when required
- Parallelization splits focused sub-tasks concurrently
- Give agents abstract tools, not specialized
Important to learn · exam keywords
EditGlobGrepReadglob patterns
Train on rails for known stops, hiker improvises anywhere
Final Assessment
- 38 questions, all topics, under 30 minutes
- Minimum request: key, model, messages, max tokens
- Lost context: Claude forgets previous messages
- Aim 80 percent before real exam
Important to learn · exam keywords
stop_reason/memoryMessage Batches APIblocking workflowdirect execution
A timed mock course before the real driving test
Module 3
Claude on Google Cloud
Vertex AI Access
- Application-default login, no API key
- Route requests through your server
- Temperature shifts sampling probabilities
- Prefill fence, stop sequence closes
Important to learn
SecurityAutomatic credentialsTemperature is a probability, not a guarantee
Librarian who checked your ID once lets you fetch freely
Prompting & Evals
- Engineering writes, evaluation measures objectively
- Examples demonstrate; start one-shot
- Score matters only when refining raises
- Combine code and model graders
Important to learn · exam keywords
edge cases
Stopwatch on every run reveals real sprinter progress
Tool Use & RAG
- Tools fetch external data
- Resend tool_use blocks and schema
- RAG chunks docs, retrieves relevant
- Hybrid search: embeddings plus BM25
Important to learn · exam keywords
tool_usetool results in conversation historyvalid empty result
Chef bookmarks only pages matching tonight's dish
Features on Vertex
- Extended thinking; max_tokens exceeds budget
- Citations return cited_text and locations
- Prompt caching reuses preprocessing five minutes
- One character change invalidates cache
Important to learn
Budget headroomCache is fragileWhen to use citations
Kitchen keeps chopped vegetables ready for repeat orders
Agents & Platform
- Workflows predetermined; agents plan themselves
- Favor workflows, resort to agents
- Give abstract tools, not specialized
- AnthropicVertex client, IAM, @-versioned models
Important to learn · exam keywords
.claude/commands/glob patternsscoped tool accesstool_choice: any~/.claude/commands/
Fixed-route bus versus taxi choosing its own path
Module 4
Claude Code in Action
Steering & CLAUDE.md
- Plan mode researches read-only before changes
- Compact summarizes context, risks dropping detail
- Worktrees isolate parallel sessions, auto-removed
- CLAUDE.md is guidance; leaner gets followed
Important to learn · exam keywords
/compactplan mode.claude/commands/independent review sessionpermission error
Map the road trip before leaving the driveway
Verification & Permissions
- Verification skill fires on matching task
- Done means gates run and observed
- Permission modes decide once, shift-tab cycles
- Auto classifier guards intent, not correctness
Important to learn · exam keywords
.claude/skills/
Lifeguard watches recklessness, not your swimming stroke
Hooks
- Hooks are code, real enforcement
- Exit 2 blocks; exit 1 does not
- updatedInput redacts secret, command still runs
- SessionStart compact matcher preserves continuity
Important to learn · exam keywords
PostToolUse
Turnstile stays locked until you tap a valid ticket
Routines & GitHub Actions
- Routine bundles prompt, repo, connectors, zero hosting
- Headless -p skips auto-discovery, faster startup
- Code Review comments; Action pushes commits
- Routines preview, hourly, claude/ branches only
Important to learn · exam keywords
plan mode--output-format json--resume.claude/skills/fresh session + structured summary
Coffee timer brews every morning without you
Trust & Plugins
- Verify in proportion to supervision
- Read the diff, not the summary
- Plugin bundles skills, hooks, MCP configs
- Plugins run with your privileges, check first
Important to learn · exam keywords
subagentcommunity MCP server
Read the label before eating a stranger's jar
Module 5
Claude 101
What is Claude
- Guided by helpful, harmless, honest principles
- Steerable thinking partner, not chatbot
- Iterate with feedback, keep context
- First drafts are starting points
Important to learn
Claude is guided by principles (helpful, harmless, honest) and designed to be steerable and…Iterate: give specific feedback on what to change, then refine
Adjust the decorator's arrangement rather than emptying the house
Desktop Modes & Results
- Three modes: Chat, Cowork, Claude Code
- Chat when answer changes next question
- Cowork for clear deliverables
- 4D: Delegation, Description, Discernment, Diligence
Important to learn
When the answer changes what you ask next (brainstorming, thinking out loud)Delegation, Description, Discernment, Diligence
Chat while sketching, hand over fixed blueprints
Projects, Artifacts & Skills
- Projects give persistent memory, context
- Artifacts are standalone shareable outputs
- Skills package repeatable expertise
- Together an organized workspace
Important to learn
Ongoing work with reference materials, consistent instructions, and team collaborationA standalone, shareable output that Claude creates (documents, code, visualizations)
Workshop holds tools, notes, and finished pieces
Connectors & Research
- Connectors integrate individual external tools
- Enterprise Search queries all sources simultaneously
- Research does multi-source investigation with citations
Important to learn
Connectors integrate individual tools; Enterprise Search queries across all connected…When you need systematic investigation across multiple sources with citations
One doorway, one all-rooms query, one reporting scout
Use Cases & Platforms
- Beyond claude.ai: Code, @Claude, Design, Office
- Same model serves every role
- Behavior shaped by context, not model
Important to learn
Claude Code, @Claude, Claude Design, Claude for Excel/PowerPoint/WordProject instructions, uploaded knowledge, and prompting style
One assistant, different notes per team
Module 6
Claude with Amazon Bedrock
Bedrock Access
- Boto3 bedrock-runtime, explicit region, converse call
- Model-not-found means region mismatch
- Cross-region inference profile routes automatically
- Bedrock stores nothing, resend alternating messages
Important to learn · exam keywords
business error
Amnesiac pen pal, retell the whole exchange each letter
Prompting & Evals
- Evals score prompts objectively across versions
- Code, model, human graders differ
- Model grader needs reasoning, else middling six
- Apply one technique, re-evaluate the gain
Important to learn · exam keywords
syntax validationsemantic validation
A stopwatch times each drill, proving real improvement
Tool Use & RAG
- stopReason tool_use, echo toolUseId back
- Resend tool schemas with results
- RAG chunks and retrieves relevant pieces
- Hybrid adds BM25, merged via RRF
Important to learn · exam keywords
end_turnMessage Batches APIRead + Write fallbackoriginal document + failed extraction + validation errorsemantic validation
Librarian pulls only needed pages, not whole book
Features on Bedrock
- Extended thinking after prompt optimization, eval budget
- Image and PDF content parts
- Citations map statements to source page
- Prompt caching, cachePoint, 1024 tokens, five minutes
Important to learn · exam keywords
explicit context passing
Underline the exact page backing every claim
Agents & Platform
- MCP servers supply ready-made tools
- Tools model, resources app, prompts user
- Computer use is just tool use
- Agents observe first, CloudWatch surfaces typos
Important to learn · exam keywords
scoped tool accessverify_fact
Hire a workshop stocked with ready-made tools
Final Assessment
- Real questions, aim eighty percent
- Tokens are smallest understood units
- MMLU tests fifty-seven subjects
Important to learn · exam keywords
few-shot examplestool boundaries
Flight simulator mirrors the cockpit before real flight
Module 7
Model Context Protocol
MCP Concepts
- Communication layer, no integration code
- Shifts tools onto specialized MCP servers
- Transport-agnostic client, usually stdio
- MCP provides schemas; tool use calls them
Important to learn
End-to-End FlowHow MCP Differs From Plain Tool UseListToolsRequest / ListToolsResult · CallToolRequest / CallToolResult
A librarian who knows the filing system fetches your book
Building a Server
- SDK generates schemas from decorators, type hints
- FastMCP server, in-memory document dictionary
- Clear description is matching criterion
- Resources expose data via ReadResourceRequest
Important to learn · exam keywords
MCP resources
A vending label describes each snack before you pick
Client & Prompts
- MCP Client class plus Client Session
- list_tools() and call_tool() core functions
- read_resource parses by MIME type
- Tools model, resources app, prompts user
Important to learn
The One-Line Mental ModelTools · Resources · Prompts
Chef picks special, host seats you, you order menu
Final Assessment
- Real course assessment questions
- Test MCP certification readiness
- Answer without looking back at lessons
Important to learn
Practice examA REST API and a GraphQL endpointCallToolRequest
Run mock driving tests before the real exam