Onboarding · Module 3

Claude Code & Cowork: How We Work

At BlueForest, Claude isn't just a chatbot — it's how we build. We work in Claude Code (the terminal agent) and Claude Cowork (the agentic app), both operating directly on local folders. This module gets you operating that way fast.

~/Recruiting — claude
$ cd ~/Operations/Recruiting
$ claude
✻ Reading CLAUDE.md… done
I now know this folder's goal, tools & conventions.

> summarize what's in this folder
✻ 2 markdown guides + an Onboarding/
  subfolder. Want me to build Module 3?
>
Part 0 · Start here

AI Fluency Check

Answer in your own words — short is fine, and "not sure" is a valid answer. Don't worry if some terms are new; that's exactly what this measures. Ammon reviews your answers, scores your fluency, and tailors your training. These submit with everything else at the final step. New to the vocabulary? The Key Concepts section right below teaches it.

A · Tooling & environment

B · Models & tokens

C · Claude Code primitives

D · Context management

E · Your current practice

F · Agentic mindset

The vocabulary

Key Concepts

Plain-English definitions of the terms above. Skim it now; come back to it as a reference.

CLI

The command-line interface — a text-based way to run programs by typing commands instead of clicking. Claude Code can run in a terminal CLI.

CLI vs Desktop app

Same Claude Code engine, two surfaces. We mostly use the Desktop app (and Cowork) — a friendlier window over the same folder-based work; the CLI is the terminal version.

Session

One continuous working conversation with Claude on a folder. It carries context until you end it or it fills up.

Token

The unit models read and write in — roughly ¾ of a word (≈ 4 characters). Usage and context are measured in tokens.

Context window

How much text (in tokens) Claude can "hold" at once. Our models run a large window — ~200K up to 1M tokens (Opus 4.8 = 1M). Large, but finite.

Opus 4.8

Anthropic's most capable Opus-tier model (claude-opus-4-8) — the strong, autonomous one we reach for on hard work. Sonnet is the balanced pick; Haiku is the fast, cheap one.

Monitoring context

Watch how full the window is (e.g. /context-style stats) so Claude doesn't lose earlier detail. Manage it by compacting, summarizing, or starting fresh.

Primitives

Skills = packaged know-how; Slash commands = shortcuts; Subagents = helpers run in parallel; Workflows = multi-agent orchestration; Hooks = automated triggers; MCP = tool connectors; Memory = what persists across sessions.

Git

Version control — saves snapshots ("commits") of a folder so you can track changes, undo, and share. GitHub hosts Git projects online.

Agentic vs chatbot

A chatbot only talks. An agent acts — it reads files, runs tools, edits, and verifies, taking multi-step actions toward a goal.

Verify / hallucination

A hallucination is a confident but wrong answer. Always check AI output (we screenshot anything visual) — trust, but verify.

Permissions

Before Claude Code runs a command or edits a file it can ask your OK. That gate is your safety net — read what it's about to do.

The core mental model

The BlueForest model: folders ARE projects

Read this first. Everything else in the module assumes it.

…/Recruiting/
CLAUDE.md
The folder's brain — goal, conventions, tools, preferences. Auto-read on open.
*.md guides
Plans, specs & notes (e.g. plan.md) — shared source of truth.
MCP tools
Workspace, Chrome, Resolve, Freepik — so Claude can act, not just talk.
git
Makes the project portable & shareable across our accounts.
  • A "project" = a local folderNot a Claude.ai "Project." A real folder on disk (often synced in Google Drive). The folder is the unit of work.
  • CLAUDE.md is the brainEvery folder has one (usually in .claude/). Code & Cowork auto-read it, so Claude shows up already knowing how we work.
  • Markdown files guide the workPlans, specs, setup guides live as .md files in the folder. You and Claude work from the same source.
  • Tools plug in via MCPWe connect Google Workspace, Chrome, DaVinci Resolve, Freepik so Claude can edit a Doc, screenshot a page, place a marker.
  • State lives in Evolving LiteNot scattered status docs. Run /evolving-lite:whats-next to see where a project stands.
  • Why this wayThe whole project lives in the folder + git — not locked inside one Claude.ai account. So we share projects across accounts and move freely between Code and Cowork. The context travels with the folder.
Just the minimum

Claude.ai fundamentals

You'll use the web/app for quick one-offs. Know the basics — then come back to folders.

  • Hold a conversationQuick questions, brainstorming, a fast image — that's what the web app is for.
  • Write a clear promptUse the pattern on the right: Context → Task → Format → Example.
  • Upload a file & ask about itDrop in a doc or image and have Claude reason over it.

The one habit to unlearn

We do not run our real work out of Claude.ai "Projects." We use folder-projects. That's the single thing to leave behind.

Prompt pattern

1
Context
What's the situation, who's it for?
2
Task
The specific thing you want done.
3
Format
How the output should be shaped.
4
Example
Show one — it anchors quality.
Part A · 2 minutes

Self-rating

Rate yourself 1–5 on each — then verify with the practical diagnostic below.

  • Terminal & file systemComfortable moving around a terminal and a file system.
  • Claude Code / CoworkHave actually used one or both.
  • Agent-directed editingCreated/edited real files by directing an AI agent — not copy-paste.
  • git / GitHubCommit, push.
  • MarkdownWritten a .md file.
  • MCPConnected a tool/connector to Claude.

Mark your honest score

Tap a bar to set each score (1–5). These submit at the final step.

Terminal & files
Code / Cowork
Agent editing
git / GitHub
Markdown
MCP
1 · New to it5 · Fluent
Part B · The real signal

Practical diagnostic

Do these in Claude Code or Cowork and save what you produce — we'll review together. Each level you clear sets your ramp path.

How leveling works

Clear the tasks at each level in order. The highest level you can do cleanly is your starting point — and it maps straight to a ramp path on the next screen.

L1
Operate in a folder → Path A
L2
Build a folder-project → Path B
L3
Plan, tool, ship → Path C / Module 4
L1

Operate in a folder

  • Open Claude on a folder and ask it to summarize what's inside.
  • Have Claude create a notes.md, then edit it on your instruction.
  • Ask Claude to organize/rename a couple of files.
L2

Build a folder-project

  • Make a new folder; have Claude write a starter CLAUDE.md (goal + conventions).
  • Scaffold something small — a one-page HTML site or a structured Google Doc.
  • Put the folder under git and make your first commit.
L3

Plan, leverage tools, ship

  • Write a short plan.md, then have Claude execute it step by step.
  • Connect a tool (Drive or Chrome) and use it — screenshot the page you built and fix what looks off.
  • Push the project to GitHub so it's shareable.
Parts C & D · Leveling → Fast-ramp

Pick your ramp — the BlueForest way

Your diagnostic level points you to a path. Each path stacks on the one before it.

Beginner · L1

Path A — Foundations

Comfortable chatting, new to folders / terminal / git.

  • Install & open Claude Code (terminal) and Cowork (app). Tip: /setup-cowork walks you through connecting tools.
  • Learn the folder-as-project loop: make a folder → add a CLAUDE.md → open Claude in it → give it work.
  • Basic terminal navigation + the prompt pattern (Context → Task → Format → Example).
Intermediate · L2

Path B — Operator

Can run Claude on a folder, make a CLAUDE.md, basic git.

  • Author a good CLAUDE.md (use the template below) — goal, conventions, tools, preferences.
  • git for sharing: init → commit → push, and why it makes a project portable across our accounts.
  • Connect MCP tools: Google Workspace + Chrome at minimum.
Advanced · L3

Path C — Builder

Plans + tools + git fluently. Go straight to Module 2 after.

  • The markdown-plan workflow: plan.md → Claude executes → you review.
  • Evolving Lite for project state (/evolving-lite:whats-next, /evolving-lite:remember).
  • Then go to Module 4 — Core Production Workflow (Remotion + Claude).
Part E · Learn by doing

Your first real task

This mirrors how we actually work. End-to-end, in one folder.

Create a new foldere.g. blueforest-practice.
Have Claude write a CLAUDE.mdStart from the template below — for "a one-page BlueForest-branded 'hello' site."
Have Claude build itLet Code/Cowork construct the site inside the folder.
Visually verify with ChromeScreenshot it, check it actually looks right, iterate. Mandatory here, not optional.
Commit to git & pushThen show Ammon the folder + the committed result.
~/blueforest-practice — claude
$ mkdir blueforest-practice && cd $_
$ claude
> write a CLAUDE.md for a one-page
  BlueForest-branded hello site
✻ Created CLAUDE.md ✓
> now build index.html
✻ Created index.html ✓
✻ Screenshotting via Chrome…
logo crisp · colors on-brand · ship it
$ git init && git commit -m "hello"
>

Visual verification is the habit

Any UI / layout / color change gets a screenshot before it ships. This is the loop we live in.

Copy into any new project folder

Starter CLAUDE.md template

The brain you drop into every folder. Fill the angle-bracket parts and Claude shows up oriented.

CLAUDE.md
# <Project Name> ## Goal <one paragraph: what this folder/project is and what "done" looks like> ## How we work here - Project state lives in Evolving Lite (/evolving-lite:whats-next) — not status docs. - Visual verification required for any UI/layout/color change (screenshot, iterate). - Ask one question at a time; keep outputs scannable. ## Tools / MCP in use <Google Workspace, Chrome, DaVinci Resolve, Freepik> ## Conventions <naming, folder structure, file/output conventions specific to this project> ## Key files <plan.md, specs, reference docs here>
  • GoalOne paragraph: what the folder is and what "done" looks like.
  • How we work hereThe conventions Claude must follow — state in Evolving Lite, visual verification, one question at a time.
  • Tools / MCPWhich connectors this project leans on.
  • Conventions & key filesNaming, structure, and the source-of-truth .md files.

For a real production example, read REMOTION-CLAUDE-SETUP.md in the QHP Remotion folder — the exact setup guide we drop into a new video project's CLAUDE.md.

Part F · Keep this handy

Cheat-sheet — Code & Cowork

Seven reflexes that make the BlueForest way automatic.

Open Claude in the folder
That's what loads the project's CLAUDE.md and files.
CLAUDE.md first
If Claude doesn't know a convention, it belongs in CLAUDE.md.
Plan in markdown
For anything multi-step, write/confirm a plan.md, then let Claude execute.
Tools via MCP
Connect Drive/Chrome/etc. so Claude can act on real things.
Verify before it ships
Always sanity-check; for anything visual, screenshot.
git = sharing + safety
Commit often; push so the project is portable across our accounts.
State in Evolving Lite
/evolving-lite:whats-next to orient, /evolving-lite:remember to capture a learning.
Submit your check-in

Send your results to Ammon

Set your ratings in Part A and tick the tasks you cleared in Part B — they're pulled in automatically. Add your details below and submit.

Check-in sent — nice work.

Ammon has your ratings, diagnostic results, and reflections. He'll confirm your ramp path. You're cleared to move toward Module 4.

Done when…

You've cleared Module 3

You can do all of this without hand-holding:

  • Open Claude Code/Cowork on a folder
  • Author a CLAUDE.md
  • Have Claude build & visually verify a small thing inside it
  • Commit it to git

Next: Module 4

Core Production Workflow — Remotion + Claude. Where the folders-as-projects model meets real video work.

All modules