System Prompt Architecture

Module 5, Lecture 5.2 | Section 3: Prompt and Context Engineering

The system prompt is not documentation — it is an engineering artifact that the agent reads on every single turn. This lecture covers how to design it deliberately: the four functional sections, why their order affects compliance, how to budget tokens, the case for native structured tool calls over text parsing, and how to write tool descriptions the model will actually follow.

Read the full lecture narrative

Additional Resources

• Lecture slides
• Tool Use with Claude — Anthropic Docs — Official overview of the native tool use API: how tools work, the agentic loop, and the token cost implications of tool definitions.
• Prompt Engineering Best Practices — Anthropic Docs — Anthropic's prompt engineering reference. Particularly relevant for its long-context guidance — placing queries at the end of a long context improves performance by up to 30%.
• Lost in the Middle: How Language Models Use Long Contexts — Liu et al., 2023 — The foundational paper behind the "lost in the middle" phenomenon. Empirical evidence that LLM performance degrades when relevant information sits in the middle of a long context — the research basis for ordering instructions at the top and bottom of system prompts.
• Writing Effective Tools for Agents — Anthropic Engineering Blog — First-party Anthropic guidance on tool description design: write descriptions for a new team member, use unambiguous parameter names, return only high-signal data, and iterate based on call metrics. Includes measured results showing concrete accuracy improvements.