Practical guides ยท Dossier 04
FORMAT: 21 PAGES, PDF PRICES FROZEN AT: 26 JULY 2026 READ TIME: ~25 MIN

Claude Code
the engineer's guide

Twenty-one pages for the engineer opening a coding agent for the first time, who would rather not spend three weeks working out on their own which habits actually matter.

Why

What's in it

The whole deck rests on one idea: Claude Code is a model called in a loop, with tools. The model picks an action, the terminal runs it, the result lands back in the context, and the loop goes again until the work is done.

What follows from that is that the quality of what you get depends as much on the harness around the model as on the model itself. Project memory in CLAUDE.md, the permission rules, tests running inside the loop so the agent catches its own mistakes before you do: those are the levers that actually move. The twenty-one pages are a long answer to the question "how do I configure that loop better", in roughly that order, with a concrete action closing every chapter.

Prices and model names are frozen at 26 July 2026, two days after Opus 5 shipped. That part will age fast: Sonnet 5's introductory price goes back up on 1 September, and whichever model is "best" has a shelf life measured in weeks. The rest of it, context management, worktrees, hooks, and how deeply to review a diff, moves far more slowly, and that's where the deck earns its keep.

What it doesn't do is replace the official documentation, which it points you at page after page. And on Agent Teams I don't have enough mileage to say anything beyond "experimental, keep review in your own hands".

The deck

The PDF reads in place below. The dense tables (pricing, slash commands, hooks) are a lot more comfortable full screen.

Mobile browsers refuse to scroll an embedded PDF, so the reader is hidden here. Open the deck full screen or download it.

Open full screen ↗ Download the PDF ↓
Contents

The six chapters

Getting started. Install, sign in, and hold a first session that's actually useful. The best opening prompt is still a question about your own code, not a request to generate something.
Models and costs. Choosing between Opus 5, Fable 5, Sonnet 5 and Haiku 4.5, and why interactive work belongs on the subscription while anything that runs without you burns API credits.
The daily fundamentals. CLAUDE.md, plan mode, the slash commands, and context management: the four habits that separate a good session from one that quietly goes sideways.
Expert level. Subagents, skills, hooks, MCP, parallel worktrees, and headless mode for CI. With the warning that belongs next to it: every automated job is code nobody reads anymore.
Pro habits. Calibrating review depth to what the code can actually break, and treating prompt injection as an architecture problem rather than something you can prompt your way out of.
The action plan. Seven days, one action each, and for every one a completion test that doesn't let you cheat.