Everything from the June 5 session, distilled to the ideas that change how you work — each with a small animation that shows the concept in motion. Read top to bottom, then test yourself on the quiz.
A normal chatbot only talks. Claude acts — it can search the web, run commands in a terminal, and create real files on your machine.
Under the hood it works in a terminal, essentially talking to itself: run a command, hit an error, fix it, run again — chaining steps until the job is actually done.
Cowork runs inside a sandbox — a sealed virtual machine on your computer. Safer for non-developers, but more restricted: it can't see folders outside the sandbox, like Downloads, Dropbox, or iCloud Drive at the top level.
Claude Code has no sandbox by default — more power, and more responsibility. It can touch anything you can.
Every Cowork or Code project is simply a folder on your computer with files inside it. Nothing magic. A fresh project starts empty — what makes it useful is what you put in it (files, and the memory below).
A CLAUDE.md file is loaded at the start of every chat in that project. It's where instructions and preferences live.
You can layer them: a global CLAUDE.md for overall preferences, plus project-level ones. They load from the folder upward — like overarching laws plus local laws.
By default a new chat does not know what happened in other chats — even inside the same project. What one chat learns stays in that chat.
That's exactly why projects matter: they keep related chats and their files together, sharing the files and the CLAUDE.md context — even though the conversations themselves don't bleed into each other.
A skill is a saved file — a prompt, sometimes with scripts or tools attached — for a repeatable task. Claude keeps a list of skills with short descriptions and auto-picks the right one when your request matches.
A plugin packages several skills and connectors together. If the "skills" section of a chat is empty, no plugin was used.
Claude decides whether to use a plugin based on its description — and it can guess wrong. You can force one via the + button or by naming it in your message.
Three models, weakest to strongest: Haiku → Sonnet → Opus. Use Opus for hard thinking (like building a skill); downgrade to Sonnet for routine tasks to save tokens.
Performance degrades as a conversation fills up — "Claude gets stupid." Under a full context, models can also start dropping steps.
On the Pro plan you get a 5-hour rolling window of tokens, plus a weekly limit. Heavy work hits these fast — another reason to use the lighter model for routine jobs.
Claude can hallucinate or stop short of finishing. Counter it with quality control — and watch out for confirmation bias: if you hand Claude your opinion, it tends to lean toward agreeing. For high-value work, consider running it once with no input from you.
A common setup scatters memory across ~6 files where the CLAUDE.md just says "always read these other files" — redundant. The simpler fix: consolidate into one CLAUDE.md where the global file belongs.
iCloud Drive ("keep downloaded") works across Macs, but the sandbox can't always reach it. Git / GitHub is the better alternative — version control plus syncing across machines, and easy collaboration.
Walk the interactive version