Hookless Architecture

Start with How it works for the user-facing mental model. This page describes the product structure behind it.

rd-skills is a non-intercepting, host-native control plane for AI engineering work:

Control Plane Prompt + four Agent Profiles + three Skill layers

It uses bounded dispatch, selectively loaded Skills, optional visible Markdown handoffs, fresh validation, and independent review when justified. It does not install executable hooks, an interception bridge, a second sandbox, private runtime evidence, or an internal task-state engine.

There is one Runtime Skill surface: 1 Control plus 25 Professional top-level Skills. Foundation and modifier-only Domain knowledge is selected JIT behind a Professional owner. Runtime composition is distinct from the four Agent Profiles shown in the architecture line above.

This shape keeps control decisions inspectable and makes host-enforcement limits explicit. It authors, builds, packages, installs, upgrades, and removes standard Skill and Profile artifacts; it is not a runtime content corpus or a personal archive system.

Continue to AI control boundaries for authority, permissions, evidence, and host enforcement. Use Operating model for routing, coordination, and completion, Subagent model for role execution, and Installation for built-artifact boundaries.