Getting Started with Tammy

Getting your own version of Tammy running takes a few hours, not a few days. The setup is front-loaded. Once the foundation is in place, adding to it is fast.

Here’s what you actually need and how it works.

Prerequisites

You need Claude Code and a Claude subscription. Claude Code is Anthropic’s official command-line tool. It installs on Mac, Windows, and Linux. The documentation is at claude.ai/code.

On subscription tier: Claude Pro ($20/month) will work to get started, but you’ll hit session limits mid-task once you start running real workloads. The Max plan is what makes this viable for sustained daily use. That’s not a pitch; it’s just the practical reality of running 26 agents against a working production schedule.

The file structure

Tammy lives in a directory on your local machine. The core files are:

CLAUDE.md is the permanent operating brief. Claude reads this at the start of every session. It contains your identity, your projects, your working style, your priorities, and your rules for how the assistant should behave. This is the document that makes the system persistent. Without it, every session starts from zero.

config/voice_profile.md is where your communication style is captured. Tone, vocabulary, banned words, how you structure arguments, how you open and close emails. The interview process builds this from samples of your actual writing, not from how you describe yourself.

todo.md is a plain text file. You say “todo: [item]” and Claude appends it with a date stamp. That’s the whole system. Simple because sessions end, but files persist.

skills/ is the directory where custom workflows live. Each skill is a markdown file with a procedure. You build skills for things you do repeatedly.

.claude/ is the configuration directory. This is where hooks, settings, and rule files live. The hooks are shell scripts that fire at session events. The session-start hook injects context. The pre-tool-use hook enforces delegation rules.

memory/MEMORY.md is the project-scoped memory index. The actual storage is in a ChromaDB MCP server with semantic search. The markdown file is a summary index.

How the interview works

The interview is how you bootstrap your CLAUDE.md and voice profile. You run Claude Code in your assistant directory and use the interview prompt to start. The system asks you nine questions:

1. Your name and role (short and long versions)
2. Your current projects and responsibilities
3. Your audience and what they care about
4. Three writing samples for voice calibration
5. What daily tools you use
6. Your top priorities for the next few weeks
7. Your working style and decision-making approach
8. Behaviors you never want to see
9. Anything else the assistant needs to know

From those answers, Claude generates two files without templates or placeholders. Your CLAUDE.md is built from your actual responses. Your voice profile is built from analysis of your writing samples, not from how you described your style.

That distinction matters. Self-descriptions of writing style are usually wrong. What you write is the truth. The interview captures the truth.

The generated outputs are: CLAUDE.md (operating brief), voice_profile.md (voice calibration), and three starter skills: distill.md, save-state.md, and catchup.md.

The three starter skills

distill runs at session end and captures stable patterns to the memory server. It identifies things worth keeping.

save-state creates a checkpoint in under 30 seconds. If you need to step away mid-session, this preserves where you were. Next session, you resume from that state instead of rebuilding context.

catchup does a light reload: reads CLAUDE.md, the time-sensitive file, the memory session start, and the todo list. Use it when context has gone stale or you’ve been away for a day.

Advanced setup (optional)

The structured memory system uses an MCP server with ChromaDB for semantic search. This requires separate setup but pays off quickly. It’s what allows Tammy to recall prior decisions, feedback patterns, and project context across hundreds of sessions.

Hooks are configured via .claude/settings.json. The four core hooks are: session-start (context injection), pre-tool-use (validation and enforcement), post-tool-use (logging), and notification (status updates). You don’t need all four to start. Session-start and pre-tool-use are the essential ones.

Starting point

Open Claude Code in your assistant directory. Use the interview prompt. Let it generate your files. Read what it produced and correct anything that doesn’t match reality.

Then use it for a week. The system reveals what it needs through use, not planning. Skills get added when you repeat something twice. Rules get tightened when something goes wrong. The CLAUDE.md evolves as your priorities change.

You don’t need everything on day one. You need the operating brief and a voice profile. Everything else grows from there.