Create rule

Create Rule

Generate a new memory rule — a project instruction, coding standard, or workflow guideline — in whatever format the project's AI tooling actually reads.

This creates memory rules, not skills or commands (use create-skill) and not a search for existing skills to install (use find-skills).

• Tool conventions, OS-specific paths, and frontmatter formats: references/memory-hierarchy.md
• Worked templates per tool and scope: references/rule-examples.md
• Claude Code memory docs: https://code.claude.com/docs/en/memory

Anti-hallucination guidelines

• Verify existing patterns first — check the actual codebase, don't assume conventions.
• Base rules on real code found in the project, not invented standards.
• If no convention exists, ask the user rather than making one up.
• Check existing CLAUDE.md / AGENTS.md / .claude/rules/*.md / .cursor/rules/*.mdc for conflicts before adding a new rule.
• Only use paths frontmatter on .claude/rules/*.md (Claude Code) or globs on .cursor/rules/*.mdc (Cursor) — never on CLAUDE.md or AGENTS.md, which have no frontmatter.

Steps

1. Parse arguments: rule name from $1 (or the first word of the arguments), description from the rest — ask the user if either is missing.
2. Detect the tool convention by checking what's already in the repo:
   • CLAUDE.md or .claude/rules/*.md present → Claude Code.
   • AGENTS.md present (with no Claude-specific files) → Codex/Cursor/Copilot/Gemini-CLI-style tool. This repo's own AGENTS.md + CLAUDE.md pair is a live example of the pattern.
   • .cursor/rules/*.mdc present → Cursor's native format.
   • More than one convention present → ask which the user wants updated (or write to more than one, if the user says so).
   • None present → ask the user which tool/format they use before writing anything.
3. Decide type and scope for the detected tool (details and templates in the reference):
   • Claude Code: modular .claude/rules/<name>.md for a focused single topic vs. a CLAUDE.md entry for a short cross-cutting instruction; project (.claude/rules/, git-shared) vs. user (~/.claude/rules/, personal) — ask if unclear from context; path-specific rules need paths frontmatter.
   • AGENTS.md-style tools: append a new section to the single root AGENTS.md — this format has no per-topic file split or frontmatter.
   • Cursor: .cursor/rules/<name>.mdc with description/globs/alwaysApply frontmatter; scope by directory nesting for path-specific rules.
4. Write the rule, following the matching template in references/rule-examples.md: imperative language, specific expectations, one topic per file, code examples where useful.

Examples

• /create-rule api-errors "Standard error handling for API routes" in a Claude Code repo → writes .claude/rules/api-errors.md with paths: src/api/**/*.ts frontmatter and concrete guidelines.
• /create-rule formatting "2-space indentation, single quotes" in a Claude Code repo, no path scope needed → writes .claude/rules/formatting.md with no frontmatter.
• /create-rule --user preferences "Prefer functional patterns, async/await over raw promises" → writes ~/.claude/rules/preferences.md (personal, not git-shared).
• /create-rule testing "Use table-driven tests and no test interdependence" in a repo with only AGENTS.md → appends a ## Testing section to AGENTS.md instead of creating a new file.
• /create-rule react-components "Function components only, props typed with interfaces" in a Cursor repo → writes .cursor/rules/react-components.mdc with globs: ["**/*.tsx"] frontmatter.