AI-DLC 2.0Workflows, explained

Home / Lesson 5 of 6

The Package Manager

The big ideaWrite each stage once in a tool-neutral file, then let the package manager emit versions for Kiro, Claude Code, Cursor and GitHub Copilot that reference it and add each tool's strengths.

Video, 1:54. On-screen captions are built in.

Teams use different AI coding tools. Without care, you would copy the same stage definition into four tool formats, and they would drift apart. The package manager prevents that with two rules:

  • Each stage lives in one canonical stage definition file.
  • Tool-specific versions refer to that file by name instead of copying its content (DRY).

But it avoids the opposite trap too: it does not shrink everything down to what every tool can do. Each tool’s emitter layers that tool’s native features on top.

Try it: one stage, four tools

Select a tool to see what its emitter generates from the same canonical stage file.

Canonical stage definitionstages/implementation.md
  • Stage structure
  • Compartment 1: inputs & outputs
  • Compartment 2: post-conditions
  • Orchestrator hints

Reference (not a copy):


      

Native affordance layered on top:

File names are illustrative. The rule from the spec: reference the canonical file by name, never copy it, and let each emitter add its tool's distinctive features.

What is shared, and what is not

In the shared definition

  • Stage structure
  • Compartment 1 and Compartment 2 contracts
  • Orchestrator hints
  • The cross-stage post-condition library

Left to each tool’s emitter

  • Claude Code skill-scoped hooks
  • Cursor Composer multi-file editing semantics
  • Copilot Workspace’s PR-native loop
  • Kiro’s spec-kit diff presentation
Try this at work

It works like a recipe card and four kitchens: one recipe, each kitchen using its own best equipment. Change the recipe once and every kitchen gets the update.

Want the original wording? Read Section 5 of the specification.

Quick check

Three questions. Get them all right to mark this lesson done.

Q1. How do tool-specific implementations use a stage definition?

Q2. Why not limit the shared definition to features every tool supports?

Q3. Which of these is part of the shared, tool-agnostic definition?