Cadence

A stack-agnostic, self-improving AI development framework as a Claude Code plugin.

Four layers (Patterns / Process / Automated gates / Launch) plus a retrospective loop that makes the framework grow as your team uses it.

Why Cadence

AI agents amplify whatever standards you give them. Without standards, they amplify chaos. With handwritten standards, they amplify yesterday’s decisions even when those have aged. Cadence is the discipline of making the standards explicit, mechanical, and self-improving.

A short tour:

Patterns (architecture.md) — your codebase conventions captured as ADRs and a PATTERNS.md
Process — Definition of Done, role specs, team protocol
Automated gates — cadence.yaml-driven verify pipeline that runs format / lint / boundary check / test from your project’s own tooling
Launch — a launch template that fills in placeholders and (optionally) spawns four coordinated subagents
Retrospective loop — every team run ends by mapping observed issues to one of the four layers, then patching that layer

Get started

Quickstart — install, init, launch, verify, retro in five minutes
Architecture — the four-layer model in detail
Self-improvement loop — how the framework grows
Cross-tool portability — non-Claude-Code use

Standards alignment

Recipes

Common stacks with worked-out cadence.yaml examples:

Case studies

Flutter sandbox — a real Flutter + Firebase project where Cadence was prototyped

Source

github.com/nik190799/cadence — Apache 2.0.