A personal AI assistant that remembers your work.

No coding. Hand one zip to your AI app (Claude Code or OpenCode) and walk away. It fits the install to your computer and, with a cloud model, finishes in under 10 minutes. Then it asks you five questions about you and how you work.

From then on it notes what you did, reads it back the next time you start, and writes you a morning briefing from it. The notes are plain files in a folder on your computer. Open them, edit them, back them up. I spent six months getting this assistant right. The zip hands it to you.

Price: $79 one-time.

Needs: a Mac, Windows or Linux computer (no graphics card needed), Claude Code or OpenCode, and an AI model. Claude Code needs a paid Claude plan or an Anthropic API account. OpenCode has free models, and a local model is free to run. See what you need.

Day one, and every day after.

Day one: your agent installs the assistant and asks who you are. Your answers become its identity files, so it knows how you like to work.

Every day after: it writes down what you worked on, reads it before you start, and briefs you in the morning once you pick a time. A health check you can schedule for each morning flags a briefing that stopped or a setting that moved.

ChatGPT and Claude remember you now too. The difference is where the memory lives. Here it is a folder you own, and you can read all of it.

Modules

Module 00: What You're Actually Building (30 min)

• Demo: assistant wakes up, reads briefing, references last week's decisions
• What separates this from a chatbot: autonomy, persistent memory, scheduled ops
• How agent skills work, SKILL.md as your assistant's identity (OpenCode or Claude Code)
• What you'll have by the end of this course

Module S: Tool Setup: Meet Your Agent (~45 min)

• Two ways to drive it: OpenCode desktop app (no terminal) or the terminal, your choice
• What an AI agent actually is: and what it isn't
• Basic usage: sessions, context, and commands
• The config file: settings that persist

Module 01: Environment Setup (~1.25 hr)

• Agent install: OpenCode desktop app or Claude Code (Mac, Windows, Linux)
• Workspace structure: ~/.openclaw/workspace/: what lives where
• Local LLM setup: Ollama (CLI) or LM Studio (GUI), same models, different experience
• API key management and rate-limit proxy, budget tracking from day one
• First run: verify your agent connects

Module 02: Your Agent's Identity (60 min)

• SKILL.md, name, description, how your agent (OpenCode or Claude Code) finds your assistant
• AGENTS.md, behavioral contract: startup sequence, tools, rules
• SOUL.md: persona, tone, model selection
• USER.md: who you are, what the agent should know about you

Module 03: Memory Without Infrastructure (65 min)

• Why every session starts blank, and three approaches that fix it without a database
• Option 1: Organized markdown, daily notes, decision log, project state. Zero setup, works today.
• Option 2: Obsidian vault, local knowledge base your agent reads and writes. Links build context over time.
• Option 3: Ripgrep search, fast text search across a notes directory. No embeddings, no config.
• How to structure notes so the agent loads the right context at session start
• When you've outgrown this: the signal that it's time for Level 2's RAG stack

Module 04: Day-to-Day Operations (115 min)

• Daily notes: YYYY-MM-DD.md, what gets written, when, by whom
• The daily briefing: yesterday's decisions, today's tasks, any blockers
• Heartbeat cron: polling on a schedule, automatic status updates
• Telegram integration: the agent reports to you, not the other way around
• Proactive monitoring: three scripts (error alerting, disk monitoring, usage reporter) with scheduled cron jobs. The system tells you when something breaks
• The morning health check: whether the scheduled briefing actually fired, not just whether it is scheduled

Module 05: Hooks: Automated Triggers (65 min)

• on-session-start: injects relevant context before you type anything
• on-session-end: writes the daily note and a structured handoff for next session
• on-compaction: archives full conversation history, then verifies nothing valuable was dropped
• prune-tool-results: stops context bloat before it slows everything down
• Telegram notify: pings you when a task completes

Module 06: Putting It All Together (2 hr)

• Start from zero, end with a working personal assistant
• Configure memory files, briefings, and three core hooks
• Run a simulated week: watch memories accumulate, see the briefing improve
• Troubleshooting: what breaks, why, and how to read errors

Module 07: Your Assistant's Graph (~45 min)

• Nodes, edges, state, and reducers, named on the assistant you already built
• The four kinds of node, and why most of yours should stay plain code
• Draw your own assistant, then read the drawing for the edges you never declared
• Declared cycles are features, discovered cycles are bugs, and every cycle needs a bound
• Why a lock file and a marker file solve different problems when your briefing runs twice

Module 08: Why You're Building This (~22 min)

• From prompts to loops: why every automation in this course is secretly a loop
• The Minimum Viable Loop: the smallest cycle worth automating
• The 4-condition checklist for whether a task deserves a loop at all
• The Loops Stack: where Level 1 ends and the Dark Factory begins

Module ★: Bonus Library ()

• The skill library: 16 production skills the Dark Factory runs on, plus a catalog indexing the 232 in the wider library
• Six specialist subagents (hephaestus, sisyphus, looper, fast-build, quality, critique) each one a single job you delegate
• Identity templates, all six course hooks, the helper scripts, and a memory quick reference
• agentvault, secure credential storage, so your agent can use your keys and never read them

What you need

The course is a one-time price. Running the assistant takes four things. Only the AI model can cost money. With OpenCode it does not have to.

1. A computer. Mac, Windows or Linux, from the last few years. No graphics card needed.
2. An AI app: Claude Code or OpenCode. Both have a desktop app, so the terminal is optional.
3. An AI model. This is the part that can cost money. Pick one:
   • With Claude Code (paid): Claude Code needs a paid Claude plan. The free Claude plan does not include Claude Code. You can also pay per use through an Anthropic API account. claude.com/pricing, code.claude.com/docs/en/setup
   • With OpenCode (free models or your own account): OpenCode is free, and it comes with free models you can use without an account. OpenCode offers them for a limited time. OpenCode says it stores your data on these models. Some are used to collect feedback. Read a model's note before you give it private work. You can also connect a provider account you already have. opencode.ai
   • With OpenCode and a local model (free): A local model is free to run. OpenCode connects to Ollama or LM Studio. You need about 8GB of RAM for a small model and 16GB of RAM for a mid-size one, and it is slower and less capable than a cloud model. With a local model and Telegram off, nothing leaves your computer. opencode.ai/docs/providers
4. About 10 minutes. With a cloud model, your agent installs Level 1 in under 10 minutes, then asks you five questions. It sees what your computer needs, like a newer Python, and sets that up as it goes. With a local model, the time depends on your computer.

Optional: Telegram, if you want the briefing on your phone. You create the bot, and your agent does the rest.

Prices and plans change. Check current prices on the official pages linked above before you pick a model. Module 2 shows you where to check your own spend.

Frequently asked questions

Why not just use free tutorials?

You could. But you'd spend months piecing together what this course hands your agent in one zip. The advice online is scattered, contradictory, and often untested. This is the collection, curated, tested, production-proven. Everything that works is here, organized and ready to use.

Do I need to be a developer?

No. Claude Code and OpenCode both have desktop apps, so the terminal is optional. Your agent writes the code. If you can install an app and follow steps in order, you can do this.

What do I need to run this?

Any Mac, Windows, or Linux computer from the last few years, no graphics card needed. An AI app: OpenCode or Claude Code, both with desktop apps; the terminal is optional. And an AI model, the one part that can cost money. Claude Code needs a paid Claude plan (the free Claude plan does not include it) or an Anthropic API account. OpenCode includes free models, offered for a limited time; OpenCode says it stores your data when you use them and uses some to collect feedback, so read a model's note before you give it private work. A local model through Ollama or LM Studio is free to run and works offline. Check current prices on claude.com/pricing and opencode.ai. Level 1 runs on all three platforms; if you continue to Level 2, its factory stack needs a Linux environment: Windows via WSL2 or 64-bit Linux (both tested; the setup script supports Ubuntu 22.04 and 24.04), or on a Mac a Linux VM or a small VPS.

How is this different from using ChatGPT or Claude.ai?

Both remember you now, and that is useful. You manage that memory through their settings. Your Level 1 assistant keeps decisions, project state, and preferences in dated notes on your computer. You can open, edit, and back up every one. Switch models tomorrow. Your context comes with you.

Does this require a server?

No server required. Everything runs on your laptop or desktop. Local models are fully supported. Nothing leaves your machine unless you choose a hosted model (your prompts and file contents go to that provider) or turn on the Telegram notifications (the notification text goes to Telegram).

What about privacy?

Every component works with local models (Ollama or LM Studio, llama3.2:3b, qwen3:8b). Your assistant's memory, daily notes, decision logs, all of it stays on your hardware. A local model is an option, not the default: the course sets up OpenCode with a cloud model, and the local-model lesson in Module 1 is optional. Run a local model with the integrations off and nothing is sent to any external server.

How long do I have access?

Lifetime access. Once you buy, you get every update to the program for free.

What is a 'harness'?

The course's word for what you are actually building. The model (Claude, or a free local model) does the raw thinking; the harness is everything around it, identity, memory, permissions, triggers, skills. Nearly all of what makes an agent useful is harness, and here every piece of it is a file you own. That is why the assistant survives model changes: the model is one line in an identity file, swap it, and the rest of the harness stays.

What is founder's pricing?

The current prices are held for the first 50 students, in exchange for an honest review once you've built. After the first 50, pricing moves to the regular rate.

What if it doesn't work for me?

30-day guarantee. Read all 10 modules in the browser, risk-free. Not for you? Email support@darkfactory.education for a full refund. Downloading the build kits makes the purchase final. The files are the product. Full policy: /refund.