Skip to content

Learning Journal

A running log of lessons I've picked up while building AI agents with Claude Code, the Claude Agent SDK, and Model Context Protocol — what worked, what failed, and which patterns I now reach for first.

Notes, experiments, and things I've stumbled into while building AI-powered tools and workflows. Most of it comes from 2025–2026 work on the OpenClaw project and ad-hoc Claude Code sessions. If you're new, start with the Claude Code essentials guide.

Purpose

I keep this journal mostly for myself — to remember what worked, spot patterns that show up across projects, and occasionally save someone else from a pitfall I already walked into.

Recent Entries

Prompt Engineering Observations

  • Structured prompts beat lengthy prose. JSON/YAML constraints give me more reliable outputs than natural language alone.
  • Tests work well as oracles. Pinning expected behaviour up front keeps the model from drifting.
  • Small, verifiable steps beat one big monolithic ask.

Tool Integration Lessons

  • File-based state is underrated. External memory (files, git) is more reliable than trusting the context window.
  • MCP makes context sharing much less painful by standardising how agents read external state.

Multi-Agent Patterns

Field notes that go alongside the broader multi-agent system overview:

  • One coordinator plus a few focused specialists is the setup I keep coming back to.
  • Shared state needs discipline. Without a structured format, agents either duplicate work or step on each other.

What I'm Exploring

  • [ ] Claude Agent SDK for production workflows
  • [ ] Eval-driven skill development
  • [ ] Long-running agent harnesses with checkpoint/resume