Claude as EA

Chatbot ~ Answers questions. Forgets everything between sessions. Needs to be told context every time.

EA ~ Manages context, drafts in your voice, surfaces what you missed, and gets better the more you use it.

Triage your inbox ~ Read every email, categorize by urgency, draft replies in your voice.

Manage your calendar ~ Check availability, create events, send scheduling links.

Draft client comms ~ Proposals, follow-ups, onboarding emails ~ in your tone.

Research and summarize ~ Pull information from docs, emails, and the web into a clean brief.

Run your code ops ~ Read files, write code, run tests, commit changes ~ from inside VS Code.

Track your business ~ Query your database, pull Stripe data, summarize pipeline status.

File access ~ Claude Code reads and writes any file in your project. No copy-paste.

Terminal control ~ Run commands, see output, iterate ~ all in the same session.

Persistent context ~ CLAUDE.md loads at the start of every session. Your EA already knows your project.

Diff review ~ Every file change shows as a diff before it applies. You approve, Claude executes.

Git integration ~ Commit, push, create PRs, resolve conflicts ~ Claude handles the ceremony.

Pro / Max subscription ~ Browser login. Fixed cost ($20 / $100~200 per month). Designed for daily heavy use. Right default for an EA setup.

Anthropic API key ~ Pay-per-token from console.anthropic.com. Best for building products, not for daily EA workflows.

Who you are ~ Name, role, location, what you do. Claude needs to know who it is working for.

Your businesses ~ Priority order, what each one does, what stage it is in.

Communication style ~ Tone, length preference, phrases to avoid, how you sign off emails.

Tools and accounts ~ What systems you use. Claude should know your stack.

Priorities ~ What you want help with most. Where to focus when you give a vague ask.

Rules ~ What your EA should never do. Guardrails that protect your brand and voice.

~/.claude/CLAUDE.md ~ Global brief. Applies to every Claude Code session regardless of project. Your identity, businesses, communication style.

project/CLAUDE.md ~ Project-specific brief. Tech stack, conventions, commands. Overrides the global file for project-specific rules.

History ~ Every change to CLAUDE.md, every new skill, every hook ~ tracked. You can see exactly when your EA learned what.

Rollback ~ Tried a new system prompt that made things worse? `git checkout` and you're back to last week's working version in one command.

Branches for experiments ~ Test a new tone or skill on a branch. Merge if it works. Throw it away if it doesn't. Your main config stays clean.

Multi-machine sync ~ Clone the repo on a new laptop and run two symlink commands ~ your full EA is configured. No re-doing setup.

Team distribution ~ Want your team to share a base config? Push to a shared private repo. Each member forks for personal additions.

Disaster recovery ~ Lose your laptop. Get a new one. Clone, symlink, work. Your operating system survives the hardware.

Commit ~ CLAUDE.md, settings.json (without secrets), skills/, hooks/, voice samples (your call), README.

Never commit ~ API keys, OAuth tokens, .env files, credential JSONs, session caches, or anything you'd be embarrassed to leak.

Borderline ~ MCP server configs that reference env vars are safe ~ the env vars themselves are not. Reference, don't embed.

Server definition ~ A config in settings.json tells Claude Code which servers to connect to and what credentials to use.

Tool registration ~ Each server registers tools ~ 'read_emails', 'create_event', 'post_message'. Claude picks the right one.

Live data ~ When you ask 'what's in my inbox?', Claude calls the Gmail server and reads your actual emails in real time.

Email summary ~ Top 5 unread sorted by urgency. Flag anything that needs a same-day reply.

Calendar ~ What is on today. Prep notes for any meetings Claude can research.

Open loops ~ Anything you flagged in the last session as a follow-up.

One priority ~ The single most important thing to do today based on context.

Step 1 ~ Triage ~ Claude reads every unread. Sorts into: reply-needed / FYI / newsletter / spam. Flags anything urgent or time-sensitive.

Step 2 ~ Draft ~ For reply-needed emails, Claude drafts a reply in your voice. Short, clear, action-forward.

Step 3 ~ Review ~ You scan the drafts. Edit what needs editing. The bar is quick iteration, not perfect prose.

Step 4 ~ Archive ~ Send the replies, archive or label the rest. Inbox to zero.