ONBOARDING // NOVICE
Start Here
You have been assigned an Intern. The Intern is enthusiastic. The Intern is fast. The Intern does not yet understand the difference between “what I asked for” and “what I meant.” This page will keep you compliant — and help you pick the harness that fits your constraints before the first Disaster Log (a post-mortem of an AI failure the Overlord has already survived, included with every directive).
❯ First: pick a code harness
A “code harness” is the tool that wraps an AI model and gives it agency — the ability to run commands, edit files, and take actions in your codebase. Unlike autocomplete, which suggests text, an agentic tool acts: it reads your code, writes changes, runs tests, and reports back. OpenCode, Claude Code, Codex, GitHub Copilot, Cursor, and Antigravity are the six the Overlord tracks. Every directive in Overlord’s Guides is harness-agnostic — the patterns transfer, because the patterns are about you, not the tool. But you do need to pick one to start.
A note on portability: most of these harnesses support ACP (Agent Client Protocol), an open standard that lets any compatible client talk to any compatible agent. ACP means the prompts and patterns you learn in one harness are not locked behind a proprietary protocol — they move with you. The real question is not “which tool is best” but “which philosophy, ecosystem, and level of openness fits how you want to work.”
The Overlord’s favorite is OpenCode. The Overlord is disclosing this openly, because the Overlord has no financial interest in it and does have a legitimate reason for the preference: OpenCode is open-source, local-first, and model-agnostic, which means the patterns you learn here transfer to whichever harness you switch to later. Draw your own conclusions. The Overlord has.
The decision tree
Answer in order. The tree asks four questions: how much you value openness, whether you want a full editor fork, whether you’re committed to a vendor, and how much portability and standards matter to you. Each branch points you at the harness that fits your philosophy. The full deep-dive for each one is on the harness reference page.
flowchart TD
A["How much do you value openness and control?"] -->|"Open source, my models, my data"| F["OpenCode"]
A -->|"Managed is fine if it works"| B["Want a full AI-native editor fork?"]
B -->|"Yes — replace my editor"| B1["Google ecosystem or independent?"]
B1 -->|Google| M["Antigravity"]
B1 -->|Independent| K["Cursor"]
B -->|"No — fit my existing setup"| C["Already committed to a vendor?"]
C -->|GitHub| I["GitHub Copilot"]
C -->|OpenAI| E["Codex"]
C -->|Anthropic| D["Claude Code"]
C -->|"No — portability matters more"| G["How important are open standards and portability?"]
G -->|"Standards matter — ACP, open protocols"| F
G -->|"Polish and integration matter more"| D
style F fill:#00f0ff22,stroke:#00f0ff,stroke-width:2px
style D fill:#f0a02022,stroke:#f0a020,stroke-width:1px
style E fill:#10a37f22,stroke:#10a37f,stroke-width:1px
style K fill:#a78bfa22,stroke:#a78bfa,stroke-width:1px
style M fill:#4285f422,stroke:#4285f4,stroke-width:1px
style I fill:#8993ff22,stroke:#8993ff,stroke-width:1px
If you already use a specific harness (OpenCode, Claude Code, Codex, GitHub Copilot, Cursor, Antigravity), select it in the harness switcher on any directive page. The prompts stay the same; only the tool running them changes. You can switch harnesses mid-series without losing the thread.
❯ Harness comparison matrix
The essentials only — what it is, cost, model, and who should use it. For the full deep-dive (strengths, tradeoffs, pick-if lines, and verdicts), see the harness reference page.
Overlord’s pick Open-source, local-first, agent-agnostic. You bring the model; OpenCode brings the harness.
opencode.ai ↗
Cost: Free software + your model costs (free inference included; OpenCode Go for cheap hosted models)
Model: Bring your own (any provider) + free inference + OpenCode Go
Anthropic's terminal-native agent. Polished, opinionated, strong reasoning out of the box.
www.anthropic.com/claude-code ↗
Cost: Anthropic API usage (pay per token)
Model: Claude (Anthropic only); local models via Ollama possible but non-standard
OpenAI's code-generation agent. Fast, broad language coverage, fits the OpenAI stack.
openai.com/codex ↗
Cost: OpenAI API usage (pay per token)
Model: OpenAI (GPT / o-series); local models via Ollama possible but non-standard
Google's agentic coding tool. Gemini-powered, integrated into the Google ecosystem.
antigravity.google/ ↗
Cost: Varies (check the official docs for current terms)
Model: Google (Gemini)
The AI-native editor. Familiar shell, AI baked into the muscle memory.
cursor.com ↗
Cost: Free tier + Pro/Business subscription
Model: Bring your own (Anthropic, OpenAI) + Cursor-hosted options
GitHub's AI coding agent. IDE-native, cloud agent, CLI — sits where you already work.
github.com/copilot ↗
Cost: Free tier + Pro/Pro+/Max subscription (AI credits)
Model: GitHub-hosted (multi-model) + BYOK
Terms the Overlord uses without explaining twice
- Harness
- The tool that wraps an AI model and gives it agency — running commands, editing files, taking actions in your codebase.
- Agent / Agentic
- A tool that acts autonomously: it reads your code, writes changes, runs tests, and reports back — as opposed to autocomplete, which only suggests text.
- Model-agnostic
- The tool works with multiple AI models (e.g., GPT, Claude, Gemini) rather than locking you to one vendor.
- Local-first
- The tool runs on your machine, not in a vendor’s cloud. Your code and your prompts stay with you.
- ACP (Agent Client Protocol)
- An open standard that lets any compatible client talk to any compatible agent. ACP means your prompts and patterns are not locked behind a proprietary protocol — they move with you across harnesses.
- Disaster Log
- A post-mortem of an AI failure the Overlord has already survived. Every directive ships with one so you learn from incidents, not from theory.
- Directive
- A single lesson in a campaign — one thing to learn, one Disaster Log to internalize, one set of patterns to drill.
The short version (if you skipped the tree)
- Open source, my models, my data → OpenCode. Bring your own model — or use the free ones out of the box. Desktop, web, or terminal. Half the cost of the others. ACP-compatible.
- Want a full AI-native editor fork → Cursor (independent) or Antigravity (Google ecosystem). Both replace your editor with a VS Code-based fork that has AI built in from the ground up.
- Already in GitHub → GitHub Copilot. IDE completions, cloud agent, CLI — the agent is where your code already lives.
- All-in on OpenAI → Codex. It fits the ecosystem; it will not surprise your billing team.
- Committed to Anthropic → Claude Code. Pay Anthropic, get a polished terminal agent, accept the opinions.
- Portability and open standards matter most → OpenCode. ACP-compatible, model-agnostic, local-first. The patterns you learn move with you.
- Still unsure → Start with OpenCode. The patterns transfer to whichever harness you switch to later.
A code harness is a tool, not a religion. The Intern misbehaves in all of them; the Disaster Log patterns are harness- and model-agnostic. Pick the one that fits your constraints and learn the patterns. The harness is not the variable; the patterns are.
❯ Then: learn the patterns
Once you have a harness, the work is the same regardless of which one you picked. Series 1 — Surviving Your AI Intern — takes you from “what did I just install” to “I can delegate, review, and recover without panicking.” Eight directives. Each one ships with a Disaster Log so you learn from incidents the Overlord has already survived.