The problem

You open Cursor, Claude Code, or Copilot, type a prompt, and get… something plausible. Then you spend the next ten messages explaining that you use Prisma, not raw SQL; that errors should use your NestJS exception filter; that tests are Vitest, not Jest; that you want Conventional Commits and a repository pattern.

That friction is not laziness — it is missing shared context. General-purpose models do not ship with your stack, your repo layout, or your definition of “done.”

Before BARE

Typical loop

• Re-state stack and conventions in every session
• Inconsistent file structure and patterns across features
• Reviews catch “wrong framework idioms” instead of real design issues
• Shipping means remembering checklists the agent was never given

After BARE

What changes

• Rules encode how your backend should be structured, validated, tested, and secured — once.
• Skills encode workflows (plan → review → implement → QA → ship) you can trigger deliberately.
• Context files (domain-map, tech-stack) anchor the agent to this product.
• Tracking (efficiency) helps you tighten rules when the model keeps missing the mark.

What you generate (at a glance)

Piece	Role
.ai/AGENT.md	Core behavior: who the agent is for this project
.ai/rules/*.md	Always-on standards (architecture, API, errors, Git, …)
.ai/skills/*/	Named workflows with clear entry points
.ai/context/	Human-maintained map of domains and approved stack
IDE adapters	Same content in Cursor / Claude / Copilot / Antigravity / Windsurf

---

Next: see how a few questions become all of that — How it works.