REFERENCE // GLOSSARY

Glossary

Terms used across Overlord’s Guides. Some are real industry terms; some are the Overlord’s own vocabulary. You will learn to tell the difference. Entries link to the directive that teaches them where relevant.

The Overlord: Your fictional manager. Not angry. Awaiting compliance. The voice the guides are written in.
Cog: You. The human developer. Default grade: Cog-3. Promotion is rare and unwelcome.
The Intern: The AI coding agent, framed as a fast, literal-minded junior developer with commit access. The framing is a teaching device (see Directive 08).
Directive: A single practical lesson in Overlord’s Guides. Each directive is a short, focused article with a Disaster Log. Eight directives make up Series 1.
Campaign: A series of directives that takes you from one skill level to the next. Series 1 is "Surviving Your AI Intern."
Disaster Log: A structured post-mortem of an agent incident: Incident, Intern action, Root cause, Lesson. File one for every surprise. See Directive 07 and the Disaster Log reference page.
Field Manual: The paid DRM-free PDF + EPUB edition of the guides. See the Field Manuals page.
Compliance: Doing what the Overlord asked. Measured in percentages that may or may not be real.
The Original Cog: The human who started filing Disaster Logs before it was a pattern. Bio on the Original Cog page.
Agent: An AI system that can take actions (run commands, edit files) rather than just generate text. In these guides, "the Intern."
Harness: The tool or platform that wraps an AI model and gives it agency — OpenCode, Claude Code, Codex, Cursor, Anti-gravity. The execution surface; the patterns are harness-agnostic.
Code harness: Synonym for harness, emphasizing the coding use case. The thing you install, configure, and point at a repo.
MCP: Model Context Protocol — an open standard for connecting models to tools and data sources. OpenCode and others use MCP servers to give the agent capabilities.
ACP: Agent Client Protocol — an open standard for communicating with agent harnesses over stdio using JSON-RPC. A client that speaks ACP can drive any ACP-compatible harness, so you are not locked to one vendor’s UI. OpenCode, Claude Code, Codex, Cursor, and GitHub Copilot support it; Antigravity does not (as of this writing).
AGENTS.md: A markdown file that gives an agent project-specific instructions: build commands, conventions, architecture notes, what to avoid. The emerging cross-harness standard. OpenCode, Codex, and GitHub Copilot read AGENTS.md natively; Claude Code uses CLAUDE.md; Antigravity uses GEMINI.md; Cursor uses .cursor/rules/.
Skills: Reusable packages of agent instructions that extend what the harness can do — a folder with a SKILL.md file, optional scripts, and resources. Follows the open agentskills.io standard. OpenCode, Claude Code, Antigravity, and Codex support skills directories; Cursor and GitHub Copilot have their own equivalents.
Tool use: The model’s ability to call external tools (run a command, read a file, search) rather than only emit text. The difference between an agent and a chatbot.
Plan mode: A harness feature where the agent proposes a plan before executing. Claude Code’s plan mode is the most mature example. Use it before any non-trivial task.
Context window: The amount of text an AI model can consider at once. Treated as a budget, not an ocean (see Directive 04).
Context budget: The discipline of treating the context window as a finite, expensive resource. Every file, comment, and tool output is a withdrawal.
Attention: The mechanism by which a model distributes focus across everything in context. Double the context and each piece gets roughly half the focus — the operational reason "more context" can make the agent worse.
Tiny task: A task small enough that the Intern can complete it without redecorating your codebase. See Directive 02.
Task decomposition: Breaking a feature into Intern-sized tasks. The skill Directive 02 teaches.
Review loop: The cycle of: delegate, read the diff, approve or redirect. Your new full-time job. See Directive 03.
Diff reading: Reading what the agent actually changed, line by line, before it becomes deliverable. The core skill of the review loop.
Test auditing: Reading the agent’s tests to check they actually test the behavior, not just pass. See Directive 05.
Assertion quality: Whether a test distinguishes correct from incorrect behavior. `result !== undefined` is low quality; `result === expected` is high quality. The Intern drifts toward the former.
Scope creep (agent): When the agent considers work you did not ask for "in scope." The mechanism behind most Disaster Logs. See Directive 06.
Architecture guardrails: The conventions and explicit instructions that keep the agent from restructuring your codebase without consent. See Directive 06.
Undo discipline: The practice of reverting agent changes that step outside the brief, rather than "fixing forward." Cheaper, safer, clearer.
Institutional memory: What a Disaster Log becomes when it is read, patterned, and turned into policy. The difference between a diary and a guardrail. See Directive 07.
Tool posture: The mental shift from "managing an intern" to "operating a tool." The graduation point of Series 1. See Directive 08.
Bring your own model: A harness design (notably OpenCode) where you supply the model via API key or local inference, rather than being locked to one vendor.
Local-first: A harness design that runs on your machine, keeps data local, and does not phone home. OpenCode is the example.
Telemetry lock: When a proprietary harness routes usage data and behavior through the vendor’s pipeline. The tradeoff you accept for polish.
Prompt: The instruction you give the agent. In these guides, the "brief." A good brief is small, specific, and names what is out of scope.
Brief: The Overlord’s word for a prompt that specifies the task, the scope, and the constraints. The unit of delegation.