Skip to main content

Module 1: Claude Code locally

Every ADW task, no matter how it's eventually run, starts from the same place: an engineer who knows how to run Claude Code on their own machine. This module gets you there.

Subscription vs. metered API — know the difference now​

This distinction matters for the rest of the track, so get it straight early:

  • Your Claude Code subscription is flat-rate. When you run Claude Code locally on your own machine against your own login, you are not paying per token. This is the unmetered local loop, and it's where every team should start and do the majority of their day-to-day agentic work.
  • The metered Anthropic API is billed per token, tracked at the organization level. This is what cloud ADW execution uses (covered in Module 6), and it is the dominant driver of the org's real LLM spend — not a rounding error.

You'll see this split again in Module 8, where it becomes a governance conversation. For now, the only thing to internalize is: running Claude Code locally on your subscription costs you nothing extra per run. That's exactly why this whole track teaches the local loop first.

Why this ordering matters

If you only remember one practical takeaway from this module, make it this: default to local. Reach for cloud execution only when a task genuinely needs to run unattended or at a scale your laptop can't give it.

Installing Claude Code​

  1. Install the Claude Code CLI following your organization's standard setup (ask in your team channel if you don't already have credentials — this track assumes you have a working login, not that you configure SSO from scratch).
  2. Confirm you're authenticated against your own subscription, not a service account or shared API key. Local ADW work should run as you.
  3. Clone (or navigate to) a repository you're allowed to experiment in.

.env and per-repo configuration​

Most Wasteology repos keep local configuration — API keys for non-Anthropic integrations, feature flags, local paths — in a .env file at the repo root, which is git-ignored. A few things worth knowing before your first session:

  • .env is for that repo's local configuration. It is not where your Claude Code subscription auth lives — that's handled by your Claude Code login itself.
  • If a repo's CLAUDE.md references environment variables (for example, a database connection string or a third-party API key), check for a .env.example or similar template first rather than guessing values.
  • Never commit a populated .env. If you're not sure whether a repo ignores it, check .gitignore before your first commit.

Your first session​

A Claude Code session in a repo with a CLAUDE.md will typically read that file automatically to pick up project-specific instructions — coding conventions, required pre-work (like the lat.md checks you'll learn about in Module 2), and any repo-specific workflow. Don't skip reading it yourself the first time, too; it tells you what the agent is about to do before it does it.

A good first session is a small, low-stakes one:

  • Ask Claude Code to explain what a specific file or function does.
  • Ask it to find where a particular behavior is implemented.
  • Make one trivial, reversible change (a comment, a typo fix, a docstring) and watch it propose the diff before anything is applied.

You are not running the full ADW loop yet — that's Module 4. Right now the goal is just comfort: starting a session, reading its output, and understanding that nothing happens to your repo without your visibility into it.

Wastey says

Your first session doesn't need to be impressive. The muscle you're building here — reading what an agent proposes before it lands — is the same muscle the human merge gate exercises later, just at a smaller scale.

Hands-on lab​

  1. Start a Claude Code session in a repository you have access to (a scratch repo is fine if you don't want to touch a real one yet).
  2. Ask it a read-only question first: "What does this repo do, at a high level?" or "Where is <some function you know exists> defined?" Confirm the answer matches what you already know about the repo.
  3. Make one trivial change through the session — a comment or a small docstring edit — and read the diff it proposes before accepting it.
  4. Note whether the repo has a CLAUDE.md. If it does, skim it and note one instruction in it you didn't expect (a required pre-task step, a forbidden action, etc.).

Practice: your first session​

your machine — flat-rate subscription, not the metered API
$ claude▋

Reveal: subscription vs. metered API​

▸Why is a local Claude Code session free per run, while the same work in cloud ADW execution isn't?

Knowledge check​

When you run Claude Code locally against your own subscription, how are you billed?
Why does this track teach the local loop before cloud execution?
What should you check before your first Claude Code session in an unfamiliar repo?

Next up: Module 2 — lat.md: grounding your agent.