AWS Learning Material Certification Cheatsheets
AIP-C01 Domain 1

Task 1.1: Analyze requirements and design GenAI solutions

Foundation Model Integration, Data Management, and Compliance · 3,235 words · source: Certified-Generative-AI-Developer-Professional-AIP-C01/domain-1/task-1-1-cheat-sheet.md

Domain 1: Foundation Model Integration, Data Management, and Compliance

Task Statement 1.1: Analyze requirements and design GenAI solutions

Domain 1 is 31% of scored exam content. Task 1.1 is the architecture and requirements-analysis foundation for the AIP-C01 exam: identify the business problem, choose a fitting GenAI pattern, validate feasibility with a proof of concept, and standardize components so the solution can move toward production.

Use this sheet to answer scenario questions that ask for the best architecture, the first implementation step, the right PoC scope, or the most reusable production-ready component.

Official Study Path

Use these in order:

  1. AWS Skill Builder: Exam Prep Plan: AWS Certified Generative AI Developer - Professional (AIP-C01)
  2. AWS Skill Builder: Domain 1 Review: AWS Certified Generative AI Developer - Professional
  3. AWS Skill Builder: Domain 1 Practice and AWS SimuLearn scenario practice
  4. AWS Exam Guide: Domain 1, Task Statement 1.1
  5. AWS Docs: Amazon Bedrock, Amazon Bedrock evaluations, AWS Well-Architected Generative AI Lens, and AWS Prescriptive Guidance for production GenAI architectures

Exam Objective Map

You should be able to:

Official skill What to know for the exam
Skill 1.1.1: Create comprehensive architectural designs aligned with business needs and technical constraints Translate requirements into a GenAI architecture pattern, select appropriate FMs and integration points, and account for latency, cost, scale, data, security, and deployment constraints
Skill 1.1.2: Develop technical proof-of-concept implementations Build a small, measurable PoC with Amazon Bedrock to validate feasibility, quality, latency, cost, and business value before full deployment
Skill 1.1.3: Create standardized technical components Use repeatable building blocks, AWS Well-Architected Framework, and the Generative AI Lens to make designs consistent across teams and environments

One-Minute Mental Model

Task 1.1 is about deciding what kind of GenAI system you are building before you build it.

Business goal
  -> user workflow
  -> data and context needs
  -> model capability needs
  -> architecture pattern
  -> PoC success criteria
  -> standardized production components

Exam shortcut:

Need to prove value first -> PoC with Amazon Bedrock and measurable criteria.
Need company facts or citations -> RAG / Bedrock Knowledge Bases.
Need actions or tool/API calls -> agentic architecture.
Need consistent multi-team reuse -> standardized components + Well-Architected review.
Need production scale -> modular services, observability, security, cost controls.

What Task 1.1 Is Really Testing

Task 1.1 is not asking you to train models from scratch. It is testing whether you can design a production-oriented GenAI solution around foundation models.

Exam area What that means in practice
Business alignment Define the user problem, expected outcome, measurable value, and risk tolerance
Technical constraints Identify latency, throughput, cost, data residency, privacy, integration, and operational constraints
Architecture pattern Choose direct FM invocation, RAG, agentic workflow, multimodal processing, batch processing, or a hybrid pattern
PoC validation Validate feasibility with a narrow implementation and explicit success criteria
Standardization Reuse model access, prompts, evaluation, guardrails, logging, deployment, and data-ingestion components
Production readiness Apply Well-Architected thinking across security, reliability, operations, performance, cost, and sustainability

Requirement Intake Checklist

Before choosing an architecture, collect requirements in these buckets.

Requirement type Questions to ask Design impact
Business outcome What decision, task, or workflow improves? What is the expected value? Defines success metrics and PoC scope
User experience Is this chat, document processing, search, workflow automation, or embedded app functionality? Drives API, UI, latency, streaming, and state design
Input modality Text, image, document, audio, video, tabular data, or mixed inputs? Determines model capability and data preprocessing
Output modality Text, structured JSON, summary, classification, code, image, action, or recommendation? Determines prompt design, validation, post-processing
Knowledge need Is general FM knowledge enough, or must answers use private/current data? General prompt vs RAG vs fine-tuning vs agent tools
Latency and throughput Interactive, async, batch, high-volume, or back-office? ConverseStream, async processing, Step Functions, queues, scaling choices
Security and compliance PII, regulated data, tenant boundaries, auditability, data residency? IAM, KMS, VPC/private access, logging, guardrails, access filtering
Integration What systems, APIs, databases, or workflows must connect? API Gateway, Lambda, Step Functions, EventBridge, agents, MCP/tools
Reliability What happens if the model, vector store, or downstream API fails? Retry, fallback, graceful degradation, circuit breaker, human review
Cost What is the acceptable cost per request, user, document, or workflow? Smaller models, prompt size control, caching, batching, routing, quotas
Evaluation What makes an answer good enough? Test dataset, automatic metrics, LLM-as-judge, human review, acceptance thresholds

Architecture Pattern Selection

Requirement clue Best-fit pattern AWS services and features to recognize
"Summarize this text", "classify this message", "generate a response" with no external knowledge Direct FM invocation Amazon Bedrock, Converse API, ConverseStream for streaming chat
"Answer from company documents", "cite sources", "use latest policy", "avoid hallucinations from stale knowledge" RAG Amazon Bedrock Knowledge Bases, embeddings, vector store, OpenSearch Service, Aurora PostgreSQL/pgvector, S3 data sources
"Process uploaded claims/forms/images and extract fields before summarizing" Multistep document or multimodal pipeline Amazon Bedrock multimodal models, Amazon Textract, Amazon Comprehend, Lambda, Step Functions, S3
"Call APIs", "look up order status", "create ticket", "take actions" Agentic workflow Amazon Bedrock Agents / AgentCore, action groups, Lambda tools, API Gateway, Step Functions
"Many teams need to use different FMs without app rewrites" Model abstraction / AI gateway API Gateway, Lambda, AppConfig, Bedrock Converse API, centralized auth/logging/routing
"High-volume offline analysis" Batch or async GenAI pipeline S3, SQS, EventBridge, Step Functions, Lambda, ECS/Fargate, Bedrock batch workflows where supported
"Human approval is required before final action" Human-in-the-loop workflow Step Functions, Amazon Augmented AI where applicable, approval queues, audit logs
"Prototype must prove feasibility before funding production" Narrow PoC Amazon Bedrock playground/API, small dataset, evaluation jobs, CloudWatch metrics, Cost Explorer

Exam trap: RAG is not the same as fine-tuning. RAG retrieves external context at request time. Fine-tuning changes model behavior by training a private copy or customized model. For current proprietary knowledge, RAG is usually the first answer.

Core AWS Services for Task 1.1

Service or feature Exam-level role
Amazon Bedrock Fully managed service for accessing, evaluating, customizing, and integrating FMs into GenAI applications
Bedrock Converse API Consistent API for conversational model invocation across supported Bedrock models
Bedrock Knowledge Bases Managed RAG capability for ingestion, embeddings, retrieval, prompt augmentation, and citations
Bedrock Evaluations Evaluate models, knowledge bases, and RAG sources with automatic, human, or LLM-as-judge methods
Bedrock Guardrails Apply content safety and responsible AI controls to model inputs and outputs
Bedrock Agents / AgentCore Build agentic applications that use tools, APIs, memory, runtime, identity, and observability
Amazon S3 Store documents, prompts, datasets, artifacts, logs, and knowledge base source data
AWS Lambda Lightweight glue for preprocessing, post-processing, tool actions, validation, routing, and API integration
AWS Step Functions Orchestrate multistep GenAI workflows, retries, approvals, branching, and error handling
Amazon API Gateway / AWS AppSync Expose GenAI capabilities to applications through managed APIs
Amazon DynamoDB Store session state, metadata, conversation history, request status, or workflow state
Amazon OpenSearch Service / Serverless Vector and hybrid search for RAG and semantic retrieval
Amazon Aurora PostgreSQL / RDS PostgreSQL with pgvector Relational application data plus vector search when SQL compatibility matters
AWS AppConfig Externalize model choice, prompt versions, routing rules, and feature flags
Amazon Cognito / IAM / IAM Identity Center Authenticate users and authorize access to apps, APIs, tools, and AWS resources
AWS KMS / Secrets Manager Encrypt data and manage credentials or third-party API secrets
CloudWatch / CloudTrail / X-Ray Monitor latency, errors, token usage, traces, audit events, and operational health
AWS CloudFormation / AWS CDK Standardize and repeat infrastructure deployment across environments
AWS Well-Architected Tool + Generative AI Lens Review and standardize architecture decisions for GenAI workloads

Skill 1.1.1: Designing the Architecture

The exam often gives you a use case and asks for the most appropriate architecture. Work from requirements to pattern.

Design dimension What to choose
Model access Use Amazon Bedrock for managed FM access; prefer a model abstraction layer when multiple models or providers may be used
API style Use Converse API for a consistent Bedrock message format; use streaming for interactive chat UX
Knowledge grounding Use RAG when answers need proprietary, current, or cited information
Workflow orchestration Use Step Functions for explicit multistep flows, retries, branching, and human approvals
Tool use Use agents or controlled tool-calling patterns when the system must take actions
Prompt governance Use versioned prompt templates and central prompt management instead of hardcoded prompts
Guardrails Apply input/output safety controls early for sensitive or user-facing apps
Data security Encrypt data, restrict access with IAM, preserve tenant boundaries, and audit access
Observability Log requests, latency, errors, token usage, retrieval quality, and model/version metadata
Deployment Use IaC, environment promotion, rollback plans, and separate dev/test/prod configurations

Common Design Flows

Direct model call:

Client -> API Gateway/AppSync -> Lambda/service -> Bedrock Converse API -> response

RAG:

Source data -> ingestion/chunking -> embeddings -> vector store
User query -> retrieve relevant chunks -> augment prompt -> Bedrock FM -> grounded answer

Agentic workflow:

User goal -> agent/orchestrator -> FM reasoning -> tool/API calls -> validated action -> final response

Production workflow:

Frontend/API -> auth -> model gateway -> prompt/template -> retrieval/tools -> guardrails
  -> Bedrock model -> output validation -> logging/metrics -> response

Skill 1.1.2: Proof of Concept Strategy

A PoC should answer: "Can this solve the business problem well enough to justify production investment?"

PoC step What to do Exam signal
Define hypothesis State what the GenAI feature should improve "Validate business value"
Select narrow use case Choose one workflow, dataset, or user journey "Before full-scale deployment"
Create representative dataset Include realistic prompts, documents, edge cases, and expected outputs "Custom evaluation data"
Test candidate FMs Compare quality, latency, cost, modality support, context length, and limitations "Feasibility and performance characteristics"
Build minimal integration Use Bedrock APIs, Lambda, S3, and Step Functions only as needed "Technical proof of concept"
Evaluate outputs Use automatic metrics, human reviewers, or LLM-as-judge where appropriate "Measure model effectiveness"
Validate safety and security Test PII handling, harmful content, prompt injection, access boundaries, and logging "Responsible production readiness"
Estimate cost and scale Measure token use, request volume, latency, and infrastructure costs "Performance and business value"
Decide next step Promote, revise, pivot, or stop based on evidence "Proceeding to full-scale deployment"

PoC Success Criteria

Good PoC criteria are measurable:

Criterion Example
Quality At least 90% of summaries accepted by reviewers
Grounding At least 95% of answers cite correct source documents
Latency P95 response time under 3 seconds for interactive flow
Cost Average cost under target per request or per document
Safety No high-severity policy violations in evaluation set
Business value Reduces manual review time by a defined percentage
Operability Logs include prompt version, model ID, latency, token usage, and error class

Exam trap: a PoC is not just a demo. The professional-level answer should include measurable evaluation, business value, operational signals, and a path to production.

Skill 1.1.3: Standardized Technical Components

Standardized components reduce repeated work and inconsistent controls across GenAI projects.

Reusable component Why it matters
Model abstraction layer Switch models, route requests, centralize retries, manage quotas, avoid provider lock-in
Prompt template registry Version prompts, parameterize inputs, support approvals, reduce hardcoding
Evaluation harness Reuse datasets, metrics, regression tests, human review workflows, and reports
Data ingestion pipeline Standardize document loading, cleaning, chunking, embedding, metadata, and sync
Retrieval service Provide a consistent API for semantic search, filtering, reranking, and citations
Guardrail layer Apply consistent safety, content, PII, and responsible AI controls
Tool/action interface Standardize Lambda tools, API schemas, input validation, and permission boundaries
Observability package Emit logs, traces, metrics, token usage, model IDs, prompt versions, and costs
Security baseline IAM roles, least privilege, KMS encryption, Secrets Manager, VPC endpoints where needed
IaC module Repeat deployment with CDK or CloudFormation across dev/test/prod
Release pipeline Promote model config, prompts, code, and evaluation datasets together
Well-Architected checklist Review designs consistently with the Generative AI Lens

Exam shortcut:

One app only -> solve the use case.
Many apps or teams -> create reusable platform components.
Production readiness -> standardize security, observability, evaluation, prompts, and deployment.

Well-Architected Generative AI Lens Cues

The Generative AI Lens extends AWS Well-Architected guidance for GenAI workloads. Map scenario requirements to the six pillars.

Pillar What the exam may test
Operational excellence Version prompts, model configs, datasets, and code; monitor quality; automate lifecycle processes
Security Protect endpoints and prompts, apply least privilege, encrypt data, audit events, secure tool calls
Reliability Handle model throttling, service failures, retries, fallback, graceful degradation, and distributed workflow completion
Performance efficiency Select the right model size, optimize latency, improve retrieval, tune prompts and context size
Cost optimization Choose cost-effective models, limit unnecessary tokens, use caching/routing, optimize vector storage and workflows
Sustainability Use efficient models and serverless or right-sized infrastructure; avoid unnecessary compute and storage

Production Architecture Anti-Patterns

Anti-pattern Better answer
Hardcoding one model ID everywhere Use a model abstraction layer or configurable model routing
Hardcoding prompts in Lambda code Use versioned prompt templates or Bedrock Prompt Management
Treating a PoC demo as production-ready Add evaluation, security, observability, cost analysis, and deployment controls
One large monolithic GenAI service Break into reusable services: ingestion, retrieval, model access, orchestration, feedback/logging
No evaluation dataset Build representative prompt/document sets with expected results and edge cases
No access filtering in RAG Preserve authorization in retrieval with metadata filters and identity-aware access
No logs for model/prompt versions Log model ID, prompt version, inference settings, latency, tokens, and errors
Agent tools trust model-generated parameters Validate all tool inputs before calling downstream systems
No fallback for model failures Use retries, circuit breakers, alternate models, queues, or graceful degradation
Using a large model for every request Route simple tasks to smaller/faster models when quality is sufficient

Scenario Decision Table

If the question says... Choose...
"Build a simple insurance claims app that extracts key fields and summarizes claims" Multistep Bedrock workflow with extraction prompt, summarization prompt, S3, Lambda, and Step Functions if orchestration is needed
"Validate feasibility before investing in full deployment" PoC using Amazon Bedrock, representative data, evaluation metrics, latency/cost measurement
"Answers must come from internal policies and include sources" RAG with Bedrock Knowledge Bases or vector store retrieval plus citations
"Support multiple FMs and switch without code changes" Model abstraction layer / AI gateway using AppConfig or configuration-driven routing
"Ensure consistent implementation across many teams" Standard components, IaC modules, shared prompt/evaluation/guardrail patterns, Well-Architected review
"Application must call enterprise APIs or complete workflow actions" Agentic design with controlled tools/action groups, input validation, least privilege
"Output must be structured JSON for downstream systems" Prompt/output schema, validation Lambda, retry or repair flow for invalid output
"Sensitive customer data is used" IAM, KMS, encryption, Secrets Manager, access controls, audit logs, PII detection/redaction, guardrails
"Need lower latency for chat UX" Streaming responses, smaller model where acceptable, prompt trimming, caching, optimized retrieval
"Need resilient operation during model outage or throttling" Retries, backoff, quotas, fallback model, cross-Region inference where appropriate, circuit breaker

Exam Traps

Trap Correct thinking
"Best model" without context There is no universal best model. Match model to use case, quality, latency, cost, modality, and governance needs
"PoC" means production A PoC validates feasibility. Production needs security, observability, scaling, CI/CD, and operations
"Use GenAI for everything" Some tasks are better solved with deterministic code, search, rules, traditional ML, or human review
"RAG fixes all hallucinations" RAG improves grounding but still needs retrieval quality, prompt design, citations, evaluation, and guardrails
"Bigger model is always better" Bigger models can increase cost and latency. Use the smallest model that meets quality requirements
"Agents are just chatbots" Agents combine model reasoning with tools, APIs, memory, orchestration, and action execution
"Security starts later" Security and compliance affect architecture from the start, especially with enterprise data and tools
"Logs can contain anything" Logs may contain sensitive prompts/responses. Apply access controls, retention, encryption, and redaction where needed

Study Checklist

Before moving to Task 1.2, make sure you can:

  • Explain Domain 1 weight and Task 1.1 scope.
  • Map business requirements to direct invocation, RAG, agentic, multimodal, batch, or AI gateway patterns.
  • Design a small Amazon Bedrock PoC with measurable success criteria.
  • Name the AWS services commonly used in a Task 1.1 architecture.
  • Explain why production GenAI systems need standardized prompts, evaluations, guardrails, observability, and IaC.
  • Apply the Well-Architected Generative AI Lens pillars to architecture tradeoffs.
  • Recognize professional-level answers that include security, cost, reliability, evaluation, and deployment strategy.

Sources