Skip to content

CONCEPT Cited by 1 source

Project rules / steering files

Definition

Project rules (also "steering files", "agent rules") = Markdown or config files checked into the repository that describe architectural constraints and coding conventions the AI agent is expected to honor, and which the agent consults automatically on every prompt. Distinct from docs for humans — steering files are a first-class agent input, read and obeyed without explicit prompting.

Canonical surface: .kiro/steering/

The 2026-03-26 AWS Architecture Blog post pins the concept to Kiro's steering-file mechanism:

"Kiro supports steering files — Markdown files stored under .kiro/steering/ — that describe architectural constraints and coding conventions. For example, a rule might state that database access must go through repository classes in the infrastructure layer. The agent consults these rules automatically, reducing the need to restate constraints in every prompt and helping to keep generated code aligned with your architecture."

This is the first wiki reference that anchors the .kiro/steering/ directory + Markdown format to an AWS source.

Why it matters

  • Reduces prompt repetition. "The agent consults these rules automatically, reducing the need to restate constraints in every prompt." The user is not expected to re-paste conventions into each conversation.
  • Reduces architectural drift. "Project rules reduce architectural drift and help maintain consistency as AI agents operate more autonomously."
  • Makes intent version-controlled. The rules diff alongside code — a PR can update conventions and the agent obeys the new shape starting next prompt.

Example rule shapes

  • "Database access must go through repository classes in the infrastructure layer" (AWS 2026-03-26 example — concepts/hexagonal-architecture enforcement).
  • "All handlers must return Result<T, AppError>; never throw."
  • "New migrations go in /migrations/YYYY-MM-DD-name.sql."
  • "No direct aws-sdk imports in /domain."
  • "Use pnpm not npm for dependency installs."

Peer mechanisms in other agents

  • AGENTS.md (Cloudflare; also the convention used by this wiki's own curator agent) — repo-root Markdown file the agent reads on every session.
  • .cursorrules (Cursor) — per-repo rules file.
  • CLAUDE.md / CLAUDE.local.md (Claude Code).
  • .github/copilot-instructions.md (GitHub Copilot).

Shape convergence: all are repo-local, Markdown / config files, read automatically, encoding architectural + stylistic intent. The 2026-03-26 AWS post names the category generically as "project rules" and points at Kiro's realization.

Caveats (underspecified in the 2026-03-26 source)

  • No rule precedence / conflict resolution documented. What happens when two steering files contradict each other, or a user prompt contradicts a rule, is not stated.
  • No schema. Rules are free-form Markdown — no enforcement that they parse, no linter.
  • No precision-recall discussion. The agent's fidelity to rules ("how often does it violate a rule anyway?") is not measured.
  • No interaction with base / system prompts documented. How steering layers relative to built-in agent policy is unknown.

Pairs with

Seen in

Last updated · 200 distilled / 1,178 read