dev-method

dev-method is a lightweight, practical method for AI-assisted software development.

It is inspired by LLM Wiki and Context Engineering practices: GitHub is the source of truth and the persistent memory for AI agents across sessions, implementers, and context switches.

What it is

• A set of principles, protocols, roles, and patterns for working with AI agents on real projects.
• Lightweight by design — useful without being bureaucratic.
• Autonomy-level aware: from fully manual (Level 0) to human-gate-only for high-risk decisions (Level 3).
• Not tied to any single AI tool, provider, or IDE.

What it is not

• Not an official Karpathy method or affiliated with any external project.
• Not dependent on Alina Lavoro (Alina Lavoro is only an advanced historical example).
• Not a framework requiring runtime code, build steps, or provider credentials.

Why GitHub as source of truth

AI agents lose context between sessions. GitHub-hosted docs serve as the persistent, inspectable, version-controlled memory that any agent or human can read and verify at any time.

Autonomy levels

The method supports four autonomy levels (Level 0–3), allowing a project to start with full human oversight and progressively delegate recoverable work to agents while keeping a human gate for irreversible or high-risk actions.

Pilots

GIS Tool remains the first real pilot project using dev-method at Level 2.5 / Level 3-track.

control-plane (mrhz1973/control-plane) is the active advanced orchestration pilot where the method is evolved under real low-touch loop pressure. Reusable strategic lessons from control-plane (for example remote hash verification, rolling implementer reports, and Cursor prompt formatting) are periodically promoted back into dev-method — see patterns/operational-repo-to-method-sync.md and examples/control-plane.md.

dev-method stays a method/reference repository, not an operational or runtime repo.

Current stable release

• v0.1.1 is the current stable tag.
• v0.1.0 remains the initial baseline.
• Stable consumers should pin to tags, not main.

Roadmap

See ROADMAP.md for planned stabilization, first pilot adoption, and future template/automation tracks.

Reading order for agents

See LLMS.md.