Long-form, one agent at a time
A guide earns its length by being specific to one tool and one version, which is also why each one gets its own section instead of crowding this page.
- Free · 29 chapters The Claude Code Field Guide How to give Claude Code context worth having, when to let it run, when to stop it, and how to catch the confident mistakes before they ship. Six parts, five appendices, verified against a named version so you can judge how stale it is.
One guide today. The next agent gets a section of its own here rather than a rewrite of this page.
Measured, not asserted
Each report is a standalone page with its data and method beside its conclusions, so you can disagree with the reasoning rather than the summary.
Thirteen days with an agent
What thirteen days of working with a coding agent actually looked like, measured from the logs: turn ratios, tool usage, cost by role, and how long runs really took.
Memory systems for LLM agents
Which systems let an agent retain facts, connect them, reason over time and let memory decay. Twenty-five scored against the same criteria.
The full library, newest first, is at /research/.
Running agents without handing over the machine
Small, public, and built because I wanted them. Each one exists to move work off the laptop the agent is driving.
claude-sandbox
An MCP server on Cloudflare Workers that runs commands in ephemeral edge containers. The agent stays on your machine; what it executes has no access to your local files.
local-llm-mcp
An MCP server that delegates subtasks — second opinions, bulk summarisation, review — to a self-hosted OpenAI-compatible model server on your own hardware.
pi-extensions
An opt-in MCP compatibility bridge for the Pi coding agent, which ships no MCP support by design. Explicit because the servers it starts run with your permissions.
How this is written
Everything here is checked before it ships, and the checking is the interesting part. An earlier edition of the guide drifted far enough in four months that three of its examples were broken as printed — hook scripts that blocked silently, permissions examples that were inert config, and a gate the text promised would stop a write and which never fired.
So each guide names the version it was verified against, each report carries the data behind its numbers, and each tool is public so you can read what it actually does rather than trust a description of it.
Agents ship roughly weekly. If anything here disagrees with your terminal, believe your terminal.