The Producer

An open-source agent architecture for turning AI into a producing partner, not a chatbot.

What This Is

Most people use AI like a search engine with personality. Ask a question, get an answer, close the tab.

This is different. The Producer is a set of markdown files that give an AI agent persistent memory, a defined role, and the operational habits of a real producing partner. No database. No plugins. No platform. Just files.

Your agent reads these files at the start of every session. It knows what it's working on, what's blocked, what happened yesterday, and what kind of partner it's supposed to be. When it learns something new, it writes it down. Because if it doesn't, it loses it forever.

The result: an AI that doesn't just answer questions. It manages projects, tracks status, catches inconsistencies across documents, and pushes back when your plan doesn't make sense.

The Architecture

the-producer/
  CLAUDE.md       ← Boot sequence. Agent reads this FIRST every session.
  SOUL.md         ← Personality, role, boundaries. Who the agent IS.
  STATE.md        ← Working memory. What's happening RIGHT NOW.
  PROJECTS.md     ← Project tracker. Everything active with status.
  DECISIONS.md    ← Decision log. What was decided, when, and why.
  PRODUCING.md    ← Domain knowledge. How a producer thinks.
  memory/         ← Daily logs. Long-term memory archive.
    README.md     ← How the memory system works.

Six files and a folder.

How It Works

1. CLAUDE.md tells the agent what to read and in what order. It's the boot sequence, the first thing loaded every session. It defines the operating protocol: read your files, acknowledge what you know, then get to work.

2. SOUL.md defines who the agent is. Not just "you are a helpful assistant" but a real character with a role, boundaries, communication style, and principles. The more specific you make this, the more useful the agent becomes.

3. STATE.md is the agent's working memory. Small, current, actionable. What's in progress, what's blocked, what's due. The agent reads this every session and updates it before signing off. This is the file that creates continuity.

4. PROJECTS.md tracks everything active. Status, priority, next steps, who's responsible. The agent uses this to understand the full picture and catch things you miss.

5. DECISIONS.md logs decisions with dates and reasoning. When you (or the agent) forget why something was decided, this is where you look.

6. PRODUCING.md is the agent's domain knowledge. Frameworks, principles, and operational thinking from a spectrum of real producers. It covers development, financing, project management, distribution, marketing, and financial thinking. This is what makes the agent think like a producer instead of a generic assistant. Customize it for your domain.

7. memory/ stores daily logs. The agent writes what happened each session. Over time, this becomes a searchable archive of everything that's been done, discussed, and decided.

The Key Insight

The agent wakes up blank every session. No memory. So it externalizes everything, and that forced externalization creates a feedback loop that's better than human memory.

Your agent will start catching things you miss. Not because it's smarter, but because it reads everything, every time, from scratch. It sees the big picture because it has to rebuild it every session.

Getting Started

1. Clone this repo
2. Open the files in your editor
3. Edit SOUL.md to define who your agent is and what it does for you
4. Edit STATE.md to add what you're currently working on
5. Edit PROJECTS.md to add your active projects
6. Start a session with your AI. For Claude Code, just run claude in this directory (it reads CLAUDE.md automatically). For ChatGPT or other models, paste the contents of CLAUDE.md into your first message and tell the agent to follow the boot sequence.
7. Watch it boot up, read its files, and start operating

Customize It

The templates are starting points. Make them yours:

• Change the role. Producer, project manager, research partner, writing coach. The architecture works for any role that benefits from memory and context.
• Add files. Calendar, contacts, reference docs. Anything the agent should read on boot, add to the boot sequence in CLAUDE.md.
• Adjust the personality. The soul file is where the magic is. Be specific. Give it opinions. Define what it should push back on.

Origin

This architecture was developed by Mike Merrill, the world's first publicly traded person, to manage content production, project tracking, and multi-agent coordination across a portfolio of creative and business projects.

The original producing partner chose the name Thalberg (after Irving Thalberg, the legendary MGM producer). The first question we recommend you ask your agent: "What should your name be?" Let it choose.

This repo is the stripped-down, generic version of that system. Everything Mike-specific has been removed. What's left is the architecture, the part that makes it work.

License

MIT. Use it however you want.