← AI-Driven Development

Adoption Guide

The layers are incremental. Adopt them in order, one at a time.

Layer 0 - Code

Establish coding standards. Not documentation about coding, but enforceable rules: naming conventions, project structure, error handling patterns. make build or equivalent enforces style and structure. No warnings, no exceptions.

Layer 1 - Observability

Add distributed tracing. OpenTelemetry + whatever backend fits your stack (Aspire for .NET, Jaeger, Grafana). Every HTTP request should produce a trace with spans for DB queries, external calls, and message handlers.

Layer 2 - Documentation

Write three short documents: Problem (what you are solving), Vision (what success looks like), Scope (what is in and what is out). Keep them brief. AI loads these into context.

Layer 3 - AI Instructions

Create instruction files that load automatically when AI works with specific file types. In VS Code: .instructions.md files with applyTo patterns. In Claude Code: CLAUDE.md. In Cursor: .cursorrules.

Layer 4 - Callable Surface

Expose the system through a callable surface that AI agents can invoke: MCP server, REST API, gRPC, or CLI. An AI agent should be able to perform any user-facing action without touching internal code.

Layer 5 - Journey Verification

Design a journey, walk the system through the callable surface, analyze the trace, evaluate against human-defined criteria. Fix violations. Re-walk until zero violations remain.

How long does this take?

Layers 0-2: you may already have these. If not, a day each.

Layer 3: an hour to create initial instruction files. Refine over the first week.

Layer 4: depends on your system. If you already have a REST API, you are done. An MCP server takes a few days.

Layer 5: first journey takes an hour to design, minutes to walk, an hour to fix findings. After that, each new journey takes less time.


Full adoption guide →