T2 Personal Assistant System
One AI assistant. 26 agents. Zero code.
Tammy is a Claude-powered personal assistant with 26 agents, 25+ custom skills, lifecycle hooks, and session persistence. TAMMY calls Gemini, ChatGPT, and Perplexity as tools when the task demands it. One orchestrator, multiple AI backends, no code written by a human.
Capabilities
What T2 does
Six core capabilities that compound. Each one works on day one. They get sharper as the system learns how you work.
Persistent Memory
CLAUDE.md operating brief, voice profile, MCP-based memory with semantic search. Nothing evaporates between sessions.
Slash Commands
25+ custom skills registered as commands with tab autocomplete. Consistent execution every time.
Skill Builder
Any repeatable workflow becomes a slash command. Describe it. Claude writes the file.
Lifecycle Hooks
Four hooks fire at session events. Identity injection, file protection, memory nudge, context restore.
Multi-AI Verification
Research and first drafts can be cross-checked against Gemini, ChatGPT, or Perplexity before anything ships. More sources, fewer errors.
26-Agent Stack
Opus orchestrates. Sonnet judges. Haiku extracts. Right agent for every task. Right model for every cost.
Architecture
How Tammy orchestrates
One assistant, three model tiers, and external AI tools called on demand. The right capability for every task.
- Research, writing, OSINT, and financial analysis
- City council, charter, and municipal coverage
- Social media distribution and SEO
- Session memory, state management, and hooks
- 26-agent orchestration across Opus, Sonnet, and Haiku
- Browser automation and tool dispatch
- Gemini: video analysis, Google Workspace access, fact-checking
- ChatGPT: cross-model validation, alternative perspectives
- Perplexity: live web search with source grounding
- NotebookLM: research library and document Q&A
- Called by Tammy as tools, not running as independent agents
- Cheapest capable tool per subtask
Workflow
How work flows
From raw source to published output. Each node is a defined handoff between agents, tools, or approval gates.
By the numbers
What's under the hood
Voice calibration
57+
Banned words and phrases trained from real corrections. The system writes like Thomas, not like AI.
Agent stack
26
Agents across Opus, Sonnet, and Haiku tiers. Gemini, ChatGPT, and Perplexity available as external tools.
Lifecycle hooks
4
SessionStart, PostCompact, PreToolUse, MemoryReminder. Fire automatically. Invisible when working.
Build your own
Three steps to a working system
You don't need T2 to use this approach. The pattern works at any scale.
Step 01
Run the interview
Paste one prompt into Claude Code. Claude asks nine questions. You answer in plain language. Your CLAUDE.md is generated from your answers.
Step 02
Install the hooks
Four shell scripts. Three configuration entries. SessionStart injects context. PreToolUse protects files. The system gets guardrails.
Step 03
Claim a skill
Describe any workflow you repeat. Claude writes the skill file. Register it in CLAUDE.md. Type the command. Done.
No code required
All of this was built by describing what was needed and letting the AI write the files. The technical work is system design, not programming. Thomas Gaume is a journalist and former Navy Fire Control Technician. He does not write code. Start here.
Documentation
Ask the reference notebook
The documentation behind this system is available as a public NotebookLM. Ask it anything about building a Claude-based personal assistant: CLAUDE.md architecture, hooks, skills, agents, session memory, and Claude Code integration.
Good starting questions: "How do I structure a CLAUDE.md?" or "What hooks does the system use and what do they do?" or "How does the distill skill work?"
Open the reference notebook