Skip to content

Emergent Cursor Rules

Zakariasson: "Rules is probably the most misunderstood concept since we launched Cursor Rules."

The misconception

Cursor Directory made rule packs discoverable — "install every rule for your stack" (Next.js rules, React rules, Python rules). Users treated rules as pre-installed stack configuration.

The correction

Rules should emerge dynamically. The workflow:

  1. Let agents run against the codebase.
  2. Watch where they go off the rails (wrong schema naming, ugly UI, missed convention).
  3. That deviation is the rule. Write it down as SOP-for-agents.
  4. Feed back — agent behavior improves, deviations shift to a new frontier, new rules emerge.

Rules are the crystallized record of past agent failures, not a pre-written style guide.

Why this framing matters

  • Models are getting good enough to follow specifics. You don't need defensive rule packs; you need corrections to actual drift.
  • Over-ruling creates its own drag. Pre-installed generic rules dilute the ones that matter.
  • It's the flywheel. Every off-rails agent run becomes a durable artifact — the rule — that improves the whole fleet next time. This is how the software-factory self-improves.
  • Analogous to org onboarding docs: nobody writes them in advance; they grow from the questions new hires actually ask.

Variants across tools

Same pattern appears with other agent systems: - claude-code — CLAUDE.md / AGENTS.md files serve the same role - zed-editor — AGENTS.md format - .cursorrules — Cursor's native format

Zakariasson notes AGENTS.md is cross-harness, which is why Cursor 3 supports it alongside native rules.

Cross-references