Skip to main content
On this page

Guide

adk is opinionated, single-operator-shaped, multi-agent. This guide explains why it works the way it does, how to install it, how to configure it once, and the few concepts every skill leans on.

Start here

Read Why
Philosophy The operating principles. One page.
Installation Clone + install.sh --target <agent>.
Multi-agent setup Capability matrix for Claude / Cursor / Codex / Junie.
overrides.yaml The one config file you maintain.
Project-scoped overrides <repo>/.adk/, <repo>/ai-guidelines/, <repo>/.temp/<task-slug>/.

Concepts (in priority order)

Concept What it gives you
Question-first Every skill asks up to 3 questions before any execution. Each Q&A is training data.
Advisor strategy Plan → clarify → present options → defer → execute → validate → report. Hand-off to /adk-explain when the user is unsure.
Decision logs Append-only JSONL of every fork. Consumed by /adk-improve to refine your defaults.
Plan/Act mode --plan literally restricts the implementer to read-only tools. Tool-level enforcement, not advisor-prose.
Edit format SEARCH/REPLACE block discipline for /adk-implement. Prevents whole-file rewrites.
Hooks PreToolUse:Bash safety + PostToolUse:Edit validator + SessionStart banner. Deterministic enforcement of the constitution.