Rules, Commands or Skills? A PM's Guide to Picking the Right One
5 min readPublished By Ofer Regev
You've been using Claude or Cursor for a while. You have a long prompt for your stakeholder updates, a paragraph about your tone that you paste at the start of every chat, and a vague feeling there must be a better way.
There is. It comes in three shapes, and they get confused constantly: rules, commands and skills. The good news is that all three are just Markdown files. The only difference is when they load and who decides to use them.
The three characters
Rules
Commands
Skills
When it loads
Always on
When you call it
When it's relevant
Who triggers it
Nobody, it's just there
You
The AI
Best for
Tone, style, conventions
Repeated workflows
Complex expert work
A short way to remember it: rules shape how, commands say do this, skills provide expertise.
A rule is loaded at the start of every conversation. You never remind the AI; it simply applies. That makes it the right home for your voice, your terminology and the conventions you never want broken. The catch: every rule costs context space, so keep them short.
A command is a speed-dial contact. You type /stakeholder-update, the saved prompt fires, and the AI never decides to run it on its own. It's perfect for the workflow you keep re-typing.
A skill is a specialist consultant. The AI scans the short descriptions of all your skills, and when your task fits one, it pulls in the full playbook, which can bundle templates, scripts and reference docs. The catch: the AI decides. A vague description means the skill never triggers.
Three questions to pick one
Should this apply to every task? Then it's a rule. Brand voice and formatting belong here.
Do I want to trigger it explicitly, the same way every time? Then it's a command.
Is it a complex, multi-step expert workflow? Then it's a skill.
Three examples from a PM's week:
Brand and voice → rule. Without it, every response sounds generic and off-brand. A rule tells the AI who it is for you, not what it does.
Writing a PRD → skill. A good PRD isn't a template, it's a process. The skill bundles the template, the section guide, the tone rules and past examples, and Claude loads it when you ask for a PRD.
Stakeholder update → command. If you're re-typing the same 40-line prompt weekly, it's a command. Paste rough notes, get an exec-ready email in your team's format.
Where MCPs fit
People mix up skills and MCPs. They do different jobs: MCPs give the AI access to your systems; skills give it expertise.
Take "plan next sprint for the mobile team." The skill loads your team's sprint-planning playbook. The MCP pulls tickets, capacity and priorities from Linear, live. The skill is the how-to, the MCP is the access, and together they do real work.
How Claude actually picks a skill
Think of a library catalog. Claude reads the card (the short description) and only pulls the book (the full skill) when it needs it. Only names and descriptions load upfront, which keeps things fast.
This is why description quality is everything. Compare:
Vague: "Helps with documents."
Specific: "Extract text and tables from PDFs. Use when the user mentions PDFs or forms."
The second one triggers. The first one sits there.
The format is also portable. SKILL.md is an open standard, so the same skill works in Claude and in Cursor. In Claude you upload skills through Settings; in Cursor you drop the file into .cursor/skills, and skills show up in the / menu too, so you can trigger them manually.
Two traps
Long chats drift. Past roughly 70% of the context window, the AI starts dropping early instructions, hallucinating and slowing down. Treat 70% as the warning light: summarize the state, then continue in a fresh chat.
Skill overload. Every skill's description loads on every prompt. Twenty skills means twenty small taxes on your context, and the AI gets worse at picking the right one. Curate, don't collect. Keep the skills you use weekly and disable the rest.
Start small
Don't build all three this afternoon. Add complexity only when you watch the AI make the same mistake twice.
This week: write one rule. Three sentences about how you write and one thing you never want to see.
Next: turn your most-repeated prompt into a command.
When you're ready: build a skill for the process you'd otherwise explain to a new teammate.
