Extend with AI

NetForge's biggest idea isn't a feature — it's a shape. The whole codebase is uniform enough that an AI assistant has exactly one way to add anything, so a complete feature can land from a single prompt as idiomatic, reviewable code.

Why it works

AI assistants do their best work when there's a clear pattern to follow and a short feedback loop. NetForge is built for both:

• One shape per layer. Every backend feature is six files under Features/{Domain}/; every screen is a folder under src/pages/. There's no "where does this go?" — there's a _Template to copy and a place for everything.
• Zero hidden wiring. Slices register by reflection and routes are generated from the file tree, so the assistant never has to find and edit a central registration file (and never forgets to). Add the folder, and it's live.
• One way to do things. RFC 7807 errors, PagedResult lists, the filter pipeline, static mapping, DbContext directly — the deliberate lack of alternatives (no MediatR, no AutoMapper, no repository layer) means there's nothing to choose wrong.
• The conventions are written down — for agents. CLAUDE.md and AGENTS.md ship in the repo as the canonical "what to do," with recipes and a conventions cheat-sheet the assistant can read before it writes.

What a prompt looks like

"Add a Projects feature: a Project has a name, description, and owner. CRUD endpoints gated by projects.* permissions, a list page with search and a create/edit dialog, and make it searchable in ⌘K."

Because there's a canonical slice to copy and documented recipes for each part, that prompt produces:

• a backend slice (Features/Projects/) with endpoints, validators, EF config, permissions, and a migration,
• a typed API module and a routed list/detail screen on the frontend,
• a search provider so it shows up in the command palette.

Each piece follows the same patterns as the rest of the app, so the diff is easy to review — you're checking that it matches the house style, not deciphering a new one.

The agent context

The repo ships the context an assistant needs to stay on-pattern:

File	Role
CLAUDE.md	The canonical agent/contributor guide — architecture, conventions, recipes.
AGENTS.md	Points any assistant (Claude, Cursor, Copilot, Codex, …) at CLAUDE.md.
docs/RECIPES.md, docs/CONVENTIONS.md	The deep how-to and one-screen reference, in-repo for local tools.

It's not Claude-specific: the agents.md convention means any AI coding assistant treats the same guide as the source of truth.

Try it

Scaffold a project, open it in your assistant of choice, and ask it to add a feature. The Add a feature recipe is the same loop a human follows — which is exactly why an assistant can follow it too.