Engineering on Kingjin.iohttps://kingjinsight.github.io/tags/engineering/Recent content in Engineering on Kingjin.ioHugo -- 0.141.0en-usWed, 11 Mar 2026 00:00:00 +0000Context Engineeringhttps://kingjinsight.github.io/posts/context_engineering/Wed, 11 Mar 2026 00:00:00 +0000https://kingjinsight.github.io/posts/context_engineering/<p>We have several parts in context:</p> <ul> <li>System prompt</li> <li>User Prompt</li> <li>Dialogue history</li> <li>Memory</li> <li>Relavent information from other sources</li> <li>Tool use</li> <li>Reasoning</li> </ul> <h2 id="what-is-context-engineering">What is context engineering</h2> <p><strong>Prompt Engineering</strong> is about how to write instructions. <strong>Context Engineering</strong> is broader - it covers strategies for curating and maintaining the optimal set of tokens during inference, including everything that lands in context window beyond just the system prompt: tools, MCP servers, Skills, external data, message history and so on.</p>We have several parts in context:

  • System prompt
  • User Prompt
  • Dialogue history
  • Memory
  • Relavent information from other sources
  • Tool use
  • Reasoning

What is context engineering

Prompt Engineering is about how to write instructions. Context Engineering is broader - it covers strategies for curating and maintaining the optimal set of tokens during inference, including everything that lands in context window beyond just the system prompt: tools, MCP servers, Skills, external data, message history and so on.

System prompt

Claude’s system prompt: link

Context Retrieval

  • Just in time context strategies
  • Rather than pre-loading all relevant data up front, agents maintain lightweight identifiers(file paths, stored queries, web links) and dynamically load data into context at runtime using tools.

Long-Horizon Tasks: Three techniques

  • Compaction

    • It takes a conversation nearing the context window limit, summarizes it, and reinitializes a new context window with the summary.
    • The art lies in selecting what to keep vs. discard, overly aggressive compaction can lose subtle but critical context. Anthropic recommends starting by maximizing recall.

  • Structured Note-Taking

    • It involves the agent regularly writing notes persisted to memory outside the context window, which get pulled back in later.

  • Sub-Agent Architectures

    • Rather than one agent maintaining state across an entire project, specialized sub-agents handle focus tasks with clean context windows. The main agent coordinates with a high-level plan while sub-agents perform deep technical work. Each sub-agent may use tens of thousands of tokens in its exploration but returns only a condensed summary of 1,000–2,000 tokens to the lead agent — achieving a clear separation of concerns.

My understanding

Write good and precise prompt at beginning. Just-in-time strategies is clever, it just work like human brain, when we need a reference/tool, we take it, otherwise we place it in somewhere, rather than memorized all these resources.

]]>