Planning & Design with AI

📋 AWS AI-DLC — Structured Planning

What AI-DLC Does Well

AI-DLC (github.com/awslabs/aidlc-workflows) excels at the planning and design phases. It's a prescriptive workflow enforced by steering files — 20+ markdown rules that the AI agent must follow through mandatory phases.

AI-DLC Phase	What It Produces
Workspace Detection	Automatically identifies brownfield vs greenfield
Reverse Engineering	Semantic model of existing architecture (brownfield)
Requirements Analysis	Structured requirements from high-level goals
User Stories	Detailed stories with acceptance criteria
Application Design	Architecture decisions and component design
Functional / NFR Design	Detailed design docs with non-functional requirements
Mob Elaboration	Mandatory team review checkpoint

Key strength

For brownfield projects, AI-DLC's automatic reverse engineering stage is compelling — the agent inspects the workspace, builds a semantic model, and uses it to constrain all subsequent phases. This is particularly powerful for large legacy codebases.

When to Use AI-DLC

Your team needs a prescribed process with mandatory checkpoints
You're working on a large brownfield codebase and want automated analysis
Enterprise governance and audit trails are non-negotiable
You want the AI agent to enforce the workflow, not developers
You need formal sign-off ceremonies (Mob Elaboration, Mob Construction)

Important: AI-DLC is a process layer that wraps around AI tools — it still requires a coding tool underneath. It lists GitHub Copilot as a supported agent. The two approaches are complementary, not competing.

🔧 GitHub — Composable Planning

Spec Kit — Spec-Driven Development

GitHub's answer to structured planning. Opt-in, agent-agnostic, living artefacts.

Constitution — Define project principles in constitution.md (coding standards, architecture constraints)
Specify — /speckit.specify generates a structured spec.md from a high-level prompt
Plan — /speckit.plan produces a technical implementation plan from the spec
Tasks — Break the plan into discrete, implementable tasks
Implement — Agent works through tasks with harnesses running
Review — PR review (human + Copilot) validates against the original spec

Planning Tools You Have Now

Tool	What It Does	Example
Copilot Chat + `@workspace`	Explore architecture before writing requirements	"@workspace What auth mechanisms exist? List all auth-related files."
Mermaid Diagrams	Force structural reasoning about architecture	"Create a Mermaid C4 diagram showing the cache service architecture."
Prompt For Questions	AI interviews you to surface requirements	"I need to add multi-tenancy. Ask me 5 questions before suggesting an approach."
Group Simulation	Simulate expert review panel	"If I showed this design to a security engineer, DBA, and platform engineer — what would each say?"
ADC Records	Capture architecture decisions	Create `adc/YYYY-MM-DD--decision-name/` with research.md, plan.md, handoff.md

The Shared Insight

Both approaches agree: unstructured AI assistance makes developers slower, not faster. Whether you use AI-DLC's steering files or GitHub's Spec Kit, the principle is the same — specs before code, structured phases, human review at checkpoints. Pick the one that fits your team's culture.