Week 2 of 10

Advanced Context Engineering

Core ≈ 3 h 50 min9 videos4 criteria

Focusadvanced prompting techniques and when each applies; RePPIT (Research, Propose, Plan, Implement, Test) and spec-driven development; MCP fundamentals (servers, clients, tools, transport); designing tools for agent ergonomics.

You buildA one-page spec, one feature shipped through RePPIT, and an MCP server with 2–4 tools

Core material≈ 3 h 50 min

  1. 1
    Video/podcast1 h 33 min
    Context Engineering
    The Pragmatic Engineer with Dex Horthy

    Context, harnesses, loops, research/plan/implement workflows, compaction, and software factories. The article and transcript are useful for notes.

  2. 2
    Article25 min
    Effective context engineering for AI agents
    Anthropic

    The canonical framing of context as a finite resource: system prompts, tools, examples, retrieval, and compaction.

  3. 3
    Video29 min
    Prompting for Agents
    Anthropic

    How prompting changes when the model runs in a loop with tools.

  4. 4
    Article25 min
    Writing effective tools for agents
    Anthropic

    Tool names, descriptions, boundaries, results, and evaluation.

  5. 5
    Specification30 min for the architecture, transports, and tools sections
    Specification
    Model Context Protocol

    Read the primary source before any tutorial.

  6. 6
    Documentation
    Spec Kit
    GitHub

    And its Agentic SDD workflow (20 min) — a concrete implementation of spec-driven development.

Deeper material

Additional video track

Build

Choose one real feature for your repository. Write a one-page spec containing objective, constraints, out-of-scope items, acceptance checks, edge cases, and a test plan. Follow Research → Propose → Plan → Implement → Test without skipping a stage, and keep the artifact from each stage.

Then build one MCP server with two to four narrow tools. Exercise it in MCP Inspector first, then from your agent. Test happy paths, invalid parameters, huge results, timeouts, and permission failures. Revise each description until a fresh agent consistently selects the right tool and passes valid arguments.

Done when

0/4