Getting Started
Two things to do before the interview: install Claude Code and create the file structure. Neither takes long. If you want something running immediately, skip to the interview and come back for the structure later.
What You Need
Tammy runs on Claude Code, Anthropic's command-line tool. You'll need a Claude Pro subscription ($20/month) at minimum. Heavier use cases (agents, automation, multi-AI tools) may need the Max plan. Start small. Scale when the workload justifies it.
There's no app to install beyond Claude Code itself. You type. TAMMY types back. If you're comfortable with a text conversation, you can use this. The setup takes an afternoon. The payoff is every day after that.
Install Claude Code
Claude Code is Anthropic's official CLI. The official documentation covers installation for Mac, Windows, and Linux.
One note on plan tier: Max plan is what makes this viable for real daily use. The session limits on lower tiers will hit you mid-task. If you're serious about building something you'll actually rely on, the Max plan is the right starting point.
Create the file structure
If you use the bootstrapping prompt on the Interview page, Claude creates this structure automatically. The details below explain what each folder does.
claude/
├── assistant/
│ ├── CLAUDE.md
│ ├── todo.md
│ ├── inbox/
│ │ └── processed/
│ ├── documents/
│ ├── output/
│ ├── tasks/
│ ├── agents/
│ └── .claude/
│ ├── hooks/
│ ├── settings.json
│ └── rules/
├── config/
│ └── voice_profile.md
├── skills/
└── [Claude projects dir]/
└── memory/
└── MEMORY.mdCLAUDE.md is the permanent operating brief. Claude reads it every session. The interview generates it from your answers.
config/voice_profile.md is the voice calibration file. Lives in the config subdirectory alongside other configuration files. The assistant reads it before writing anything in your name.
todo.md is the persistent capture file. Say todo: mid-task and the item gets written here without interrupting your work. Say anything on the todo list? to review what has accumulated. See the Features page for how it works.
skills/ holds your custom workflow files. It lives alongside the assistant directory, not inside it. The three core skills (/distill, /save-state, and /catchup) are generated here during setup. You add more as your workflow grows.
.claude/ is where Claude Code looks for hooks, settings, and rules. hooks/ contains shell scripts that fire at session events. settings.json configures which hooks run and when. rules/ holds directory-specific instruction files that Claude reads when working in those directories.
memory/MEMORY.md lives in the Claude projects directory, where Claude Code stores project-scoped data. It points to the MCP memory server where Claude stores and retrieves memories using semantic search. Old topic files stay here as read-only backups.
inbox/ is the file drop zone. Documents, agendas, PDFs that need processing land here. processed/ is where the originals go after routing.
documents/ holds categorized working files. output/ holds finished deliverables. tasks/ is for inter-agent handoffs when you get into multi-agent workflows.
Optional: memory and hooks
This is not required on day one. The interview gets you running. But when you're ready to go deeper:
Structured memory lives in an MCP server backed by ChromaDB with semantic search. This requires a separate installation and configuration step: the MCP server must be set up and registered in your Claude Code settings before the memory tools are available. Once running, when Claude learns something new about your workflow, it saves a memory to the server using memory_save. At cold start, memory_session_start surfaces the most relevant memories automatically. Old markdown topic files stay in the memory/ folder as read-only archive backups, but all active memory operations go through the MCP server.
Hooks are shell scripts that fire at session events. They live in .claude/hooks/ with config in .claude/settings.json. The features page covers what each hook does and why. You don't need them day one, but they make a real difference once the assistant is calibrated.
First session
Start Claude Code in your claude/assistant/ directory. If CLAUDE.md is already populated from the interview, Claude reads it and starts calibrated. If it's blank, Claude greets you normally and waits.
The more you put in CLAUDE.md, the more useful the first session is.
Run the interview before or right after setup. There is no wrong order. Go to the interview.
Use the reference notebook while you build
A public NotebookLM contains all the T2 site documentation plus key Claude Code references. Use it as a companion resource while you set up your system. Ask it how any part of the architecture works, what belongs in a CLAUDE.md, or how to structure hooks and skills.
It answers from the source material directly, so you get specific, grounded answers rather than general AI guesses about how the system works.