A publication on context engineering

agents.cli

"Agents are capable.
They're just missing your context."

Start a course Read the foundations

The thesis, live

Same agent. Same task. Different context.

Watch one agent run the same request twice — once bare, once with the project's context mounted. Flip the switch yourself. Everything on this site is about that difference.

> Add rate limiting to the invite endpoint.
Start closing the gap: the foundations

Try the ideas

Don't read about it. Operate it.

Every core idea on this site ships as something you can poke. Three of them, live on this page.

The engine

Every agent is the same loop.

Claude Code, Codex, Copilot, Cursor — under the branding they all run one loop: gather context, act with tools, verify the result. Step through a real bug fix and watch it turn. Once you see the loop, context engineering is just feeding it better.

How agents work
Step 1 / 8You hand over a task
> The login test is failing on main. Fix it.
(prompt enters the window)

Your message lands in the context window. The model reads it and decides what it needs first — it can’t fix what it hasn’t seen, so the first move is almost never an edit.

The map

Know where every file lives.

CLAUDE.md or AGENTS.md? Committed or gitignored? Where do skills, slash commands, and hooks actually go? Browse the real config tree for each tool — every entry says what it does, when it loads, and whether it belongs in git.

The configuration chapter
your-project/
~ (home directory)
Select a file or folder

Every entry shows what the file does, when Claude Code reads it, whether it belongs in git, and a working example.

The router

Say the problem. Get the primitive.

Thirteen primitives close the gap — rules, skills, subagents, MCP, hooks, plan mode, and the rest. You don't need all of them on day one. Pick the sentence that sounds like your problem and it routes you to the chapter built for it.

All thirteen foundations
I want the agent to…

Pick the sentence that sounds like your problem — it routes to the primitive built for it, and says what not to reach for.

What you'll find here

Three surfaces. One thesis.

Agents are broad, fast, and contextless. The humans they work with are the inverse: slower and narrower, but rich in context. Everything on this site is about closing that gap.

The courses

Learn by doing, one agent at a time. Pick Claude Code or Codex and
take a real project across a week of work — from your first change to
running it headless in CI. Each lesson is a real problem and the move that fixes it.

Foundations

Thirteen techniques for closing the agent–context gap. Rules, subagents, skills, MCP, hooks, plan mode, permissions, and the rest of the toolkit.

Blog

Essays on how teams actually work with AI coding agents: what compounds, what doesn't, and what's still unsettled.