Dw System — AI skill for Claude Code

dw-system — Ship With AI, Reliably

**English** | [中文](./README_zh.md)

   

**The methodology layer your AI coding stack is missing.** Prompting harder doesn't make AI coding reliable — engineering does. dw-system decomposes a raw requirement, layer by layer, into tasks so well-specified that **a zero-context AI agent can execute them dependably** — through Triage, top-level design, engineering specs, coding plans, coding, and tests.

This is not a whiteboard framework. It is distilled from an AI-assisted development system **proven in real production across 50–60 brownfield repositories under bi-weekly, high-velocity iteration**, condensed into eight single-purpose skills that load only when the workflow reaches their step.

The whole system follows **two governing principles**, which are also what it teaches your project: **every fact has exactly one home — everywhere else only links to it**; **normative text has a size budget — when over budget, relocate first, condense second, raise the ceiling last**.

Every project that installs it maintains one **mandatory foundation**: the `engineering-context/` **code-fact layer** (a capability registry + one map card per repo + one domain card per business domain). No tiers, no switch, regardless of project size or headcount. Initialization scaffolds it; from then on the rule is *need a fact about a module? Check the card first. No card? Read the source, then write the card on the spot* — committed with the same task branch, never deferred.

On top of that foundation, exactly two things vary:

• Collaboration surface (asked once, stored in config) — is anyone e