Twelve ideas, one per screen. Show each one working, then stop and ask where it fits their work. Leave the silences in.
Four moves: understand the tool, get it running, learn the one workflow, then customise it to your work.
Claude.ai hands text back and waits. Claude Code reaches into your files, terminal and codebase and does the work itself.
An agent reads and understands the codebase, edits files across the project, runs terminal commands, and searches the web, then decides what to do next.
Everything else sits on this: prompt, gather context, act, verify, and loop back until the result holds up.
Four parts make the loop run: the loop itself, a managed context window, tools it can call, and permissions you control. That combination is what a chat window doesn't have.
The same agent runs in five places. Wherever you start it, it can see that folder and everything inside it.
Want the cutting edge? Terminal. Want it inside your editor? VS Code or JetBrains. Want it working while you do other things? Desktop. Working from a GitHub repo remotely? Web.
Press Shift + Tab to cycle how hands-on you are. There's no right answer, only what you're comfortable with.
Plan mode is the safe place to start anything complex: it reads your code, asks clarifying questions, and hands back a plan before a single file changes.
Most people jump straight to "write the code" and course-correct forever. This is the workflow that saves the rework.
Make "correct" explicit (success criteria), give it tools and a test suite to check against, and if it keeps hitting the same wall, tell it to save the fix to CLAUDE.md.
Every file read, command run and message sent takes space. Three commands keep it usable.
Spend less context: be specific (vague prompts cost more, not less), turn off MCP servers you're not using, and hand exploration to a subagent that returns just the answer.
The agent that wrote the code carries its own bias. A subagent reviews it cold, in its own context.
Keep the reviewer read-only: it flags issues, it doesn't edit. Check it into the repo so the whole team reviews the same way.
A plain file that holds the rules of your work, read at the start of every session, so Claude starts already briefed.
It's an onboarding script for your codebase, appended to every prompt. Commit it so the team shares one source of truth. Keep it compact and only what's necessary.
Key conceptThree levels, stacked: project ./CLAUDE.md (committed, team), local ./CLAUDE.local.md (gitignored, just you), user ~/.claude/CLAUDE.md (every project on your machine). → full detail in the teacher notes.
A subagent does the messy exploration in its own context and hands back only the answer. Your main context stays clean.
"Where are the auth endpoints?" The subagent reads half the codebase to find out, but only the answer comes back. The journey never clutters your main window.
A skill advertises itself with a name and one line. Claude loads the full instructions only when it decides to use it.
That's the win over loading everything up front: skills keep the context window lean. (The course points to a dedicated "Introduction to agent skills" for the deep dive.)
Most of your context lives in databases, apps and repos. MCP lets Claude Code reach it through tools.
Watch the cost: every server loads its tool definitions into context, used or not. Run /mcp and disable what you're not using. A CLI like gh, or a Skill, is often leaner.
A hook is a rule that runs automatically at fixed points in the lifecycle, every time, no exceptions.
Auto-format on every edit, log every command for compliance, block writes to production. If it needs to happen every time without fail, don't put it in a prompt; put it in a hook.
Explore → Plan → Code → Commit is the spine. Context, CLAUDE.md, subagents, skills, MCP and hooks are how you make it yours.
tutto.one/praxis · daniel@tutto.one