Claude is not a chatbot you ask questions. It is an operating capability you build around yourself. Meet the surfaces, claude.ai, desktop, mobile, and set the expectation for the whole program. Do this now: open claude.ai and start one real task you have today.

Create a Project named after a real part of your work and turn on memory. This is the home your Claude learns you in. Do this now: create one Project and enable memory.

The setup almost nobody does. Paste the Profile Builder into a fresh chat and answer honestly. It hands back a Personal Profile and Initial Instructions to save. A complete profile carries your name, a bit of history, your writing style, output preferences, no-flattery and candor rules, and how you like to be asked. Do this now: run the Profile Builder and save both blocks.

Claude does what you tell it, not what you wish. A wish, be helpful, changes nothing. A rule, answer in five bullets and never use jargon, changes the output and you can check it. Climb from wish to rule to a full system. Do this now: rewrite one wish as a rule and test it.

Bring a real task and finish it with Claude: give context, review, ask for a change, refine. Do this now: complete one real task and note how you got there.

Use Claude on real work for a week and log three tasks. This is the raw material Claude turns into a profile of you. Do this now: log your first task.

A connector gives Claude access to your apps with your own permissions, and you approve each one. Claude drafts your email but never sends it. Do this now: connect Gmail, Drive, or Calendar and run one task on your real data.

With connectors live, Research searches the web and your own inbox, calendar, and docs into one sourced answer. Do this now: run one Research task grounded in your own data.

A Skill is packaged expertise Claude loads on demand. Ask in plain language and Claude picks the right built-in Skill, a deck, a spreadsheet, a document. Do this now: build one real deliverable with a Skill.

Use the skill creator: describe your repeatable process and Claude builds the Skill. Keep it narrow. Give it your inputs, your steps, one hard rule, and an example of good output. Do this now: build one Skill from your own process and run it.

Connectors give access, Skills give the procedure, the sandbox is the engine. When something fails, name which layer broke. Do this now: identify which layer caused one failure you hit.

On the desktop app, give Claude a whole goal and let it plan, split the work, and return a finished result. Do this now: run one Cowork task end to end.

Claude Code is Claude in your terminal, IDE, or desktop. The trust dial runs from plan mode, show me first, to auto mode, run trusted actions. Do this now: run one real task in Claude Code and set sensible permissions.

A Skill is a note for one project. A Plugin is a package, skills, agents, hooks, connectors, you install from a marketplace and share with a team. Do this now: decide which of your Skills should become a shared Plugin.

MCP servers connect custom tools, subagents delegate separable work, hooks automate deterministic steps. Do this now: connect one custom tool, delegate one task, add one hook.

Schedule recurring work: a daily brief, a weekly report. Verify it actually fired, do not assume. Do this now: schedule one recurring job and confirm its output.

Choose one function to internalize first, on cost, fit, and data ownership. Design the slice: the data, the agents, the parallel-run and cutover plan. Do this now: fill a one-page slice design.

Build the slice, run it in parallel with the incumbent, migrate when the evals hold, then decommission. Never cut over on a demo. Do this now: get one function to production and plan the next.