Claude Code · AI Engineering
The AI Layer
Build the harness that makes Claude Code work in a large codebase
1
A codebase now has three parts
AI LAYER
← the new part
TESTS
CODE
Code and tests, you already know. The AI Layer is the third part: the configuration and tooling you build around the model so an agent is productive in your codebase.
Anthropic's playbook says it plainly — the harness matters as much as the model.
2
The AI Layer = seven components
CLAUDE.md hierarchy
Lean root + one file per service/package, loaded additively
CLAUDE.md ×8
Hooks
SessionStart orientation + self-improving Stop hook
.claude/hooks/
Skills
On-demand expertise, progressive disclosure
.claude/skills/
LSP
Symbol-level navigation, pyright instead of grep
docs/lsp-setup.md
MCP
AST-based structured search: where_is / find_references / outline
tooling/mcp/
Subagent
A read-only explorer that maps a subsystem
.claude/agents/
Plugin
Bundles the portable pieces into one installable package
tooling/helpline-ai-layer/
3
Take it to your own codebase
⬇ Install the plugin
One command installs hooks, subagent, MCP, generic scoped-tests skill
→ Point your agent at the repo
Reads AI-LAYER.md, builds a matching layer for your code
4
Keep the layer alive
The AI Layer is not set-and-forget. Review it every 3–6 months and after every major model release. Models evolve. Your codebase evolves. The layer needs to evolve too.