## 10.1 Prompt Engineering Fundamentals **Priority:** Should Understand **Executive Definition:** Prompt engineering is writing and structuring the instructions given directly to a model (clarity, examples, explicit format requests, role framing) to get more reliable output from a single request. It is a real, learnable skill with measurable effect on output quality, but it operates entirely within one call to the model and has no influence over what data, tools, or memory that call has access to. **Why It Matters:** Prompt engineering still matters, and labs continuously publish updated, concrete guidance because specific technique reliably improves output quality. But it is a tactical, per-request skill: it cannot fix a request that lacks the right retrieved context, tool access, or permissions: that is the system-architecture problem covered next, which is where most enterprise AI quality gains now come from. Treat this as broad organizational literacy every knowledge worker should have, not a strategic investment area in itself. **What I Need to Understand:** - Well-documented, low-cost techniques (specific instructions, examples, explicit output format, letting the model reason before answering) measurably improve output quality and cost nothing to adopt broadly. - Prompt engineering is bounded by what's in the single request: it cannot retrieve data the model wasn't given, cannot recall past sessions, and cannot call tools that weren't provided (that is Context Engineering's job: next topic). - This is a skill for broad organizational literacy (every knowledge worker benefits from basic competence), not a specialized function to scale headcount around. - Prompt-level gains are use-case-specific and often marginal at the system level compared to fixing retrieval, permissions, or tool access: don't over-invest here at the expense of context engineering. **Questions I Should Be Able to Ask My Team:** 1. Do we have a basic, shared internal guide for effective prompting so common techniques are used consistently rather than reinvented per team? 2. When a use case underperforms, have we ruled out a system/context problem (retrieval, tools, permissions) before spending more effort on prompt wording? 3. Are we distinguishing ownership/versioning of fixed prompt templates embedded in applications from ad hoc end-user prompting in our governance approach? **Technologies / Standards / Companies to Know:** Anthropic/OpenAI/Google prompting guides, few-shot prompting, chain-of-thought, prompt template/version management. **Recommended Learning:** - [Prompt engineering best practices for 2026 (Anthropic)](https://claude.com/blog/best-practices-for-prompt-engineering): current, primary-source guidance. - [Claude prompting best practices (Claude Platform Docs)](https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices): the technical reference version of the same guidance. **Time Investment:** 30 minutes --- ## 10.2 Context Engineering & System Architecture Around the Model **Priority:** Must Understand **Executive Definition:** Context engineering is the discipline of deciding what information, tools, and memory a model actually receives for a given task (the retrieved documents, conversation history, tool definitions, and system instructions assembled around it) as distinct from prompt engineering, which only shapes the wording of the instruction itself. It treats the model as one component in a larger system whose overall design determines output quality far more than any single prompt does. **Why It Matters:** This is the central point of this entire domain: as models have become broadly capable, the gap between good and poor enterprise AI outcomes has shifted from "which model" or "how the prompt is worded" to "what the system around the model actually gives it access to." Anthropic's own engineering guidance frames context engineering as the natural evolution of prompt engineering precisely because agentic and long-running tasks fail on context management (what's retrieved, what's remembered, what's pruned) not on phrasing. Every earlier topic in this domain (permissions-aware retrieval, data provenance, memory, model routing) is a context engineering concern; prompt engineering is one narrow, comparatively minor input to it. **What I Need to Understand:** - Context engineering covers everything that goes into a model's context at runtime (retrieved documents, tool outputs, conversation/memory, system instructions) curated and pruned deliberately rather than dumped in wholesale. - Long-running agentic tasks fail primarily from context problems (irrelevant or excessive information crowding out what matters, stale memory, missing tool results), not from imperfect prompt wording: this is why "just improve the prompt" stops working as systems get more agentic. - Good context engineering actively decides what to leave out as much as what to include; more context is not automatically better (see the degradation-with-length point in Context Windows). - This is where investment should concentrate: retrieval quality, permission-aware data access, memory curation, and tool design will move enterprise AI outcomes far more than incremental prompt refinement. - It requires cross-functional ownership (data, security, and application engineering together) not a single prompt author. **Questions I Should Be Able to Ask My Team:** 1. When an AI application underperforms, do we diagnose it as a context problem (wrong/missing retrieval, stale memory, missing tool access) before treating it as a prompt or model problem? 2. Who owns end-to-end context assembly for our production AI systems: is it a defined engineering responsibility, or is it happening ad hoc inside prompts? 3. What share of our AI investment is going to retrieval/permissions/memory architecture versus prompt-level work, and does that match where the evidence says the returns actually are? **Technologies / Standards / Companies to Know:** Anthropic context engineering guidance, agent frameworks (LangChain/LangGraph, LlamaIndex), Model Context Protocol (MCP), the retrieval and memory systems covered in the topics above. **Recommended Learning:** - [Effective context engineering for AI agents (Anthropic)](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents): the primary source for this topic's framing and evidence. - [Contextual Retrieval (Anthropic)](https://www.anthropic.com/engineering/contextual-retrieval): cross-reference: a concrete, foundational context-engineering technique. **Time Investment:** 2-3 hours