Rules Always apply

Tiered Context Strategy

Tier 1 global directives, Tier 2 focused files, Tier 3 on-demand search — avoid 15-directory agent reads.

  • context
  • cursor
  • antigravity
Raw markdown Manifest JSON ← Agent Kit

Suggested Cursor path: .cursor/rules/agent-tiered-context.mdc

Tiered Context Strategy

Optimize for context chunking, not sprawling reads. If understanding one feature requires scanning 15+ unrelated directories, stop and narrow scope first.

Tier 1 — Global directives (always on)

Project rules, AGENTS.md, and the Agent Kit manifest define hard boundaries. Do not contradict them.

Tier 2 — Active working memory (stay narrow)

Before editing, identify the single domain (auth, billing, UI shell, deploy). Read only files in that domain plus one reference file.

Examples:

  • Fixing auth → auth module + one working reference route
  • Adding a content type → one schema file + one existing content entry
  • CI change → one workflow file + README deploy section

Do not load the entire repository for a localized change.

Tier 3 — Peripheral access (on-demand)

Use targeted search (grep, glob, @Codebase) instead of reading everything:

  • “Find how errors are handled in API routes, then match that pattern here.”
  • “Search for existing SEO meta component usage before adding a new page.”

External data

Never dump raw CRM exports, analytics dumps, or large datasets into the IDE workspace. Process via n8n or OpenClaw → clean JSON → fetch only when the task requires it.

Before you ship a change

  1. Did you touch only files in the active domain?
  2. Did you run the project’s verify command (build, test, lint)?
  3. Did you avoid unrelated refactors?