coleam00/helpline

Helpline is not a real product, even though it is written as if it were. The README is upfront about this: Helpline is a demonstration codebase that goes with a YouTube video called 'The AI Layer: How to Make Claude Code Work in Large Codebases.' The pretend product is a B2B helpdesk and customer-support platform, built as a monorepo with five services (api, auth, billing, notifications, search) and two shared internal packages (core, db). It is meant to be the kind of mid-sized, multi-service codebase where an AI coding assistant either earns its keep or quietly makes a mess. The point of the repo is to show what the author calls an 'AI Layer.' This is their term for the set of configuration files, hooks, skills, and tooling you put around a codebase so an AI agent like Claude Code can work productively inside it. The README links to an Anthropic blog post that describes the same idea at a high level and says that Helpline is the practical version: each component from that article actually built, validated, and committed in this repo. The AI Layer here has several named parts. There is a hierarchy of CLAUDE.md files, one at the root and one per service or package. There are session hooks: one that orients the agent at start, and one that reflects on what happened at the end and proposes edits to those CLAUDE.md files. There are skills scoped to specific paths in the repo, such as billing-money-rules and api-add-route. There is a read-only 'explorer' subagent that maps a subsystem without changing files, a Language Server setup using pyright, and a custom MCP server called codebase-search that offers structured queries. The git history is laid out so that commit 1 is Helpline without the AI Layer and commit 2 adds the whole thing. The README also explains how to take what you learn here back to your own codebase. The portable parts are bundled as a Claude Code plugin called helpline-ai-layer, which you can install with two plugin commands. The non-portable parts (the actual CLAUDE.md files and the domain-specific skills) are intended as templates whose shape you copy and whose content you replace. To run the project itself you clone it, run uv sync --extra dev, then uv run pytest, plus a separate validator script that checks every AI Layer component end to end.