Guide

Rules vs. Skills vs. Prompts: which do you need?

Every instruction format answers the same two questions differently: when does it reach the model, and who decides? Pick by the job, not the name.

Rules

.mdc · .cursorrules
Reaches the model:
Automatically, every session
Who triggers it:
Nobody — always on

Policy and constraints: “never commit secrets”, “always use our error type”.

You need this when: You keep correcting the agent for the same thing.

Browse rules

Project instructions

AGENTS.md · CLAUDE.md
Reaches the model:
Automatically, in one repository
Who triggers it:
Nobody — ambient context

What this project is, its commands, conventions, and hard limits.

You need this when: Every new session re-explains your project from scratch.

Browse project instructions

Skills

SKILL.md
Reaches the model:
On demand, when the task matches
Who triggers it:
The agent

Reusable procedures: “how we write changelogs”, “how to run visual QA”.

You need this when: You paste the same how-to steps into chats repeatedly.

Browse skills

Prompts

plain text
Reaches the model:
Once, when you send it
Who triggers it:
You

One-shot behaviors: “review this diff like a staff engineer”.

You need this when: You want a behavior right now, in any tool, no setup.

Browse prompts

Rules of thumb

  • Correcting the same mistake twice? That correction is a rule.
  • Pasting the same steps twice? Those steps are a skill.
  • Onboarding the agent to the repo every session? Write AGENTS.md once.
  • One-off transformation or persona?That's a prompt — start there, and promote it to a rule or skill when you reach for it again.
  • Start small: a focused 20-line rule beats a 300-line manifesto the model skims.

Where do tools, plugins, and MCP fit?

Everything above is knowledge — text the model reads. Tools, plugins, and MCP servers are capabilities: they let an agent do new things (query a database, control a browser) rather than know new things. Waxmark catalogs the knowledge layer; capabilities have their own registries. If a “tool” you found is really a description plus steps, it probably belongs here as a skill.

Whatever you pick, it arrives sealed: how Waxmark scans every file before it reaches your agent.