AI Foundations, Prompt Engineering and Context Engineering

1. AI concepts and boundaries

• The problem with the word “AI”
  • Why AI is not a single thing
  • How marketing language can obscure capabilities and limitations.
• AI vs traditional software
  • Rules-based and model-based behaviour: determinism, probabilistic outputs and overlaps.
  • Rules, logic, workflows, and automation
• A practical view of AI approaches, not a single universal definition
  • Prediction, generation and goal-oriented behaviour in selected systems.
  • Pattern learning vs explicit instruction
• Common categories of AI use
  • Prediction and classification
  • Generation (text, images, code)
  • Decision support and optimization
  • Planning and task execution
• Limits of capability claims
  • Distinguish observable task performance from claims of human-like understanding.
  • Do not infer human intent from fluent output.
  • Assess reasoning-like outputs against evidence and task results.

2. How model-based systems work

• The basic AI system pipeline
  • Data, model design, objectives, training and inference: a simplified lifecycle.
• Learning from examples
  • Learning patterns alongside rules, objectives and other system components.
  • The role of probability in AI behavior
• Training vs using a model
  • Training, possible updates/fine-tuning and repeated inference: different lifecycle activities.
• Why AI outputs change
  • Randomness and probability
  • Context sensitivity
  • Environmental inputs and system state
• Predictability and reproducibility in model-based systems
  • Deterministic components, generation settings and statistical behaviour.

3. Ways to classify systems

• Classification by learning approach
  • Supervised learning
  • Unsupervised learning
  • Self-supervised learning
  • Reinforcement learning
• Classification by capability
  • Narrow task-specific AI
  • General-purpose models
  • Multi-modal systems
• Classification by autonomy
  • Assistive systems
  • Semi-autonomous systems
  • Tool-using systems with configured autonomy and oversight.
• Risk and impact as a contextual assessment, not a universal legal taxonomy
  • Lower-consequence productivity examples.
  • Decision-support examples with relevant risks.
  • Consequential or autonomous tasks requiring stronger safeguards.
• Why classification matters for governance, cost, and trust

4. Language models

• What an LLM actually is
  • Text prediction at scale
  • Tokens as text units: not necessarily whole words.
• How LLMs “generate” responses
  • Probability distributions over possible next tokens
  • Why fluent language does not equal correctness
• Training lifecycle
  • Pretraining on large corpora
  • Instruction tuning
  • Preference and alignment tuning
• Why LLMs sound confident even when wrong
  • Fluency bias
  • Lack of grounding
• Practical limits of LLMs
  • Context window constraints
  • Limits of information learned during training versus current retrieved sources.
  • Sensitivity to phrasing and order

5. Limitations and risks

• Non-determinism
  • Why the same input does not always produce the same output
  • Reproducibility challenges
• Black-box behavior
  • Interpretability techniques and limits; generated explanations are not proof of internal causation.
  • The danger of over-interpreting justifications
• Hallucinations
  • Fabricated or unsupported information: possible causes and verification.
  • Conditions that increase hallucination risk
• Emergent behavior
  • Unexpected capabilities
  • Apparent performance changes: check evaluation methods and evidence.
• Over-reliance and automation bias
  • Human trust failures
  • Deskilling risks
• Security risks
  • Prompt injection
  • Data leakage
  • Tool misuse in agentic systems

6. Prompt design and refinement

• What prompt engineering actually is
  • Instruction design, not magic words
• Anatomy of a good prompt
  • Task definition
  • Role specification
  • Constraints and boundaries
  • Output format control
• Prompt clarity and ambiguity management
  • Assumptions
  • Definitions
  • Scope limits
• Prompt patterns to test for the selected model and task
  • Decompose a complex task into manageable parts when useful.
  • Structured output prompts
  • Request relevant evidence, checked calculations or concise rationale rather than assume displayed reasoning proves correctness.
• Prompt anti-patterns
  • Overloaded prompts
  • Conflicting instructions
  • Implicit expectations
• Prompt lifecycle management
  • Versioning
  • Testing
  • Iterative refinement

7. Context engineering

• Why prompts alone are not enough
  • The difference between instruction and environment
• What “context” means in AI systems
  • System messages
  • Retrieved knowledge
  • Tool outputs
  • Conversation history
  • Hidden operational constraints
• Context vs memory
  • Short-term context
  • Persistent information or memory mechanisms where the application provides them.
• Designing effective context
  • Relevance filtering
  • Ordering and prioritization
  • Noise reduction
• Context failure modes
  • Context overload
  • Conflicting signals
  • Stale or misleading information
• Context engineering for non-technical roles
  • Business rules as context
  • Policies and guardrails
  • Brand/tone and applicable policy context; prompt text alone does not enforce compliance.

8. Retrieval and grounding

• Why source checking and evaluation remain necessary
• Retrieval-augmented generation (conceptually)
  • External knowledge sources
  • Use retrieved evidence as context; retrieval does not guarantee factual correctness.
• Embeddings and similarity (high-level intuition)
• Designing traceable knowledge flows
  • Source selection
  • Update frequency
  • Attribution and traceability
• Risks of poor grounding
  • Confident misinformation
  • Outdated or biased sources

9. Agentic workflows

• What agentic AI means in practice
  • Goals, plans, actions, feedback
• Difference between chatbots and agents
• Single-agent vs multi-agent systems
• Tool-using agents
  • APIs, databases, browsers, workflows
• Autonomy spectrum
  • Suggestion-only systems
  • Human-approved execution
  • Execution within configured permissions and limits, with risk-appropriate oversight.
• Risks unique to agentic AI
  • Runaway behavior
  • Compounding errors
  • Actions based on unreliable outputs or an incorrectly specified objective.
• Guardrails for agentic systems
  • Permissions
  • Limits
  • Human-in-the-loop controls

10. MCP connectivity

• The integration problem in AI systems
• Why tools, data, and models need standardized interfaces
• Conceptual overview of MCP
  • Client/server access to context resources and tools, not automatic shared memory between agents.
  • Tool access
  • Separation of responsibilities
• When MCP-style approaches make sense
• Security and governance implications
  • Trust boundaries
  • Auditability
  • Access control

11. Evaluation and monitoring

• Why conventional software tests need complementary model/task evaluations.
• Defining “good enough” behavior
• Evaluation dimensions
  • Accuracy
  • Usefulness
  • Safety
  • Consistency
• Human vs automated evaluation
• Monitoring AI in real use
  • Drift
  • Degradation
  • Misuse
• Incident response for AI failures

12. Responsibility and governance

• Ethics as a practical and organisational concern
• Core ethical principles translated into practice
  • Fairness
  • Accountability
  • Transparency
  • Privacy
• Organizational responsibility
  • Who owns AI decisions
  • Who approves deployment
• Bias and representational harm
• Legal and compliance awareness (high-level)
• Building trust without over-promising

13. Organisational design choices

• Build vs buy decisions
• Choosing the right level of AI sophistication
• Cost, risk, and value tradeoffs
• Cross-functional collaboration
  • Product
  • Legal
  • Security
  • Operations
• Change management and AI adoption

14. End-to-end review

• Framing the problem correctly
• Selecting the right AI approach
• Designing prompts, context, tools, and oversight together
• Embedding safety and evaluation from day one
• Knowing when not to use AI