Write & build

Where you write code — and the agents that write it with you.

This setup has two families of tool: editors, where you type code yourself, and AI agents, which read your project and make changes for you. Knowing what each one is — and when to reach for it — is most of the skill.

#Two families

An editor is where code lives on screen: you open files, type, and save. An agent is a program you give a task to in plain English — it reads the project, proposes changes, and applies them once you approve. You'll use both, often at the same time: an agent drafts, you refine by hand in the editor.

#The editors

#The agents, and the loop they share

The terminal agents all work the same way once you start them in a project folder. You describe a task; for anything sizeable they show a plan first; they make edits; and they show you a diff to approve before anything is kept. Learning this loop once means you can use any of them.

To start any of them, change into a project folder and run its command:

$ cd ~/code/my-project
$ claude
# then just describe what you want, in plain English

#Give agents a concrete target

The single highest-leverage change in how you direct an agent: replace a description of what you want with an artifact the agent can check itself against. An agent with a checkable target can verify its own work and iterate; an agent with only prose has to guess what you meant and wait for you to notice. Every round-trip you save compounds.

This is spec design and eval design in miniature, and on real teams the habit hardens into repo machinery — for example, a rule that every bug fix starts with a red test before any fix is written. Groundwork teaches the habit; a well-governed repo enforces it.

#The desktop AI apps

Not everything needs the terminal. These graphical apps are installed (on personal machines — work machines skip AI apps) for chat, quick questions, and supervising agents:

#Copy context for review

Sometimes the best reviewer is outside the tool that can read your repo. In that case, send a small, deliberate context bundle: the files that explain the question, each with its path attached. VS Code has built-in commands for the Command Palette and copying the active file path, but open-tab contents need an extension.

#Editor extensions stay lean

Groundwork keeps VS Code and Cursor extensions focused on the agent loop: collect context, review changes, inspect CI, run UI tests, and enforce project-local language rules. It avoids duplicate snippet packs, overlapping formatters, and helper extensions that mostly replace built-in editor behavior.

#The shared language: AGENTS.md

The agents look for plain-text project instructions and follow them. In this setup, AGENTS.md is the canonical file: it says how to run the tests, what style to use, what never to touch, and what verification means. Tool-specific files like CLAUDE.md should point back to it instead of repeating the same rules.

You've already seen this idea in action: the knowledge base uses AGENTS.md as its schema. Same mechanism — a file that turns a general agent into one that follows your rules.

#When to reach for which

#TODOs without another extension

VS Code already gives you the useful parts of a TODO list. Use Search across the workspace for TODO|FIXME|HACK|NOTE, or use ripgrep in the terminal:

$ rg -n "TODO|FIXME|HACK|NOTE"

The old Todo Tree extension is not part of this setup anymore. It duplicated built-in search and could fail when its bundled ripgrep dependency was missing.

#Good habits

#Know your AI budget

Every serious AI tool runs inside limits — a context window, plan or rate limits, sometimes a spend budget — and long or complex work fails in confusing ways when you hit them blind. The habit and where each tool's gauge lives now have their own page: Know your AI budget. Why it counts as fundamentals: you are the senior partner, and knowing the machine's limits before assigning it work is the same judgment you'd apply to any teammate. Token budgets are one half of that awareness; the machine's own CPU and memory (the callout above) are the other.

#Practice

#Going deeper

• Claude Code — Quickstart — install to first change; the best entry point for the primary agent here.
• Claude Code — Best Practices — Anthropic's patterns for working well with an agent (explore, plan, verify, scale).
• AGENTS.md — the open convention for project instructions, shared across coding agents.
• Claude Code — memory — Claude-specific behavior for project memory and file references.
• Anthropic's interactive prompt tutorial — a hands-on course on prompting well.
• OpenAI Codex docs & Grok Build — the other agents installed here.