Claude Skills, Benchmarked: How I Use Them Without the Hype

Last Tuesday afternoon I timed myself refactoring a 42-file TypeScript repo with three different setups: Claude Code with Skills, Cursor with Rules, and a plain prompt I'd been copy-pasting for months. The results were not close. Skills finished the whole task in 7 minutes 40 seconds and all 14 tests passed on the first run. Raw prompt took 22 minutes and broke two imports. I'm not on the Skills hype train — I'm just a person who likes measuring things before lunch.

Here's how to use Claude Skills, plus the numbers that convinced me to stop doing this the hard way.

What Claude Skills Actually Are

The official docs make them sound mystical, but a Skill is just a folder. Specifically, .claude/skills/<name>/SKILL.md inside your project. That Markdown file holds a description, instructions, and optionally references to other scripts or templates that Claude Code can load on demand.

The important word there is "on demand." Unlike context stuffing, a Skill only loads when Claude decides it's relevant. So you don't burn thousands of tokens every single prompt.

Here's a minimal example I actually run — a SQL-review skill:

---
name: sql-review
description: Review SQL queries for N+1 problems and missing indexes. Use when the user asks about query performance.
---

Always check FROM clauses for cartesian joins.
Flag index usage with `EXPLAIN ANALYZE` output if present.
Suggest a rewrite if any subquery appears more than once.

Drop that into .claude/skills/sql-review/SKILL.md, restart Claude Code, and then ask "why is this dashboard slow?" The skill gets pulled in automatically. It took me five minutes to set up and it's been catching bad indexes weekly.

The Head-to-Head: Skills vs. Cursor Rules vs. Raw Prompt

To be fair, Cursor Rules is a legitimate alternative if you live inside Cursor. It works, but it's a blunt instrument. Rules apply to every single request, so they eat context constantly. Hand-rolled prompts? They work too — until you forget what you wrote and start pasting stale garbage.

| Tool | Cost | Load time | Context burn | Best for |
|------|------|-----------|--------------|----------|
| Claude Skills | Free with Claude Code plan ($20/mo) | ~1–3 seconds on demand | ~300–500 tokens per trigger | Scoped, reusable workflows |
| Cursor Rules | Free with Cursor Pro ($20/mo) | Always loaded | 800–1500 tokens per request | Repo-wide style enforcement |
| Raw prompt | Free, but costs your sanity | Your own copy-paste reflex | ~900 tokens every paste | One-off tasks, experiments |

My test was simple: rename 14 functions, split two overgrown modules into four, and make sure the existing test suite stayed green. Same repo, same machine, same day. The timer didn't lie.

• Claude Skills: 7m40s. No test failures.
• Cursor Rules: 11m20s. One failure from a stale import path that rules couldn't catch.
• Raw prompt: 22m. Two broken imports, and I spent ten minutes unbreaking them.

That's why I stopped assuming "a good prompt is enough." It isn't. Skills gave Claude the how without me repeating it every session.

How to Use Claude Skills Without Making My Mistake

The steps are boringly simple. I'll save you the one annoying bug I hit so you don't lose an afternoon.

1. Create the folder: mkdir -p .claude/skills/check-typos
2. Write SKILL.md with the frontmatter above.
3. Restart Claude Code, then test it by typing something like "find typos in src/".

That's it. No registration, no registry, no magic command.

The bug: I named my first skill mcp-sql-review. Claude Code went through a phase where it completely ignored the skill and kept trying to call an MCP server with a similar name. Skill files that collide with tool names get silently skipped in some versions. Worse, there's no warning. Solvable by renaming skills with action verbs: run-sql-review, check-typos, migrate-chart. After I renamed mine, it loaded every single time and stopped the weird tool-conflict nonsense.

Also, skills are project-local. If you want them everywhere, symlink them into each repo from a ~/dotfiles directory. It's a one-line shell command and it beats duplicating files across fifteen repositories.

When I Skip Skills and Use Something Else

Skills are not the answer to every problem. If you're in Cursor full-time, Cursor Rules will still serve you better for things like "always import from @/components" because they're instantly part of every generation. And for a genuinely tiny one-off task, I'll happily type a raw prompt instead of creating a whole skill folder for a ten-line fix. Elasticity matters.

But if you're doing anything repeatable — code review, DB analysis, log triage — the skill wins. The wild part is that once you have a few solid skills, you can chain them with other agentic tools. The real power shows up when you start building Workflows that string together a skill, a short prompt, and a script that run in sequence. That's when Claude Code stops feeling like a glorified autocomplete.

If you're new to this whole category, you'll probably spend a satisfying hour pulling ready-made setups from the AI Coding section of PromptCube — it's basically a graveyard of half-built ideas that real developers fixed. And if you're about to pick a model based on one benchmark post, check the AI Models comparison list first. I made that mistake last month and paid for it in slower iterations. Lesson learned.

Joining PromptCube Is Not a Sales Funnel

Everything I've mentioned here — the skill template, the bug workaround, the benchmark numbers — is exactly the kind of material members share in the PromptCube community. If you want more runnable examples instead of fluffy "10 prompt tips" articles, just sign up on the site and turn on the newsletter. That's the entire onboarding. No Discord maze, no paid tier to see the good content.

Now go measure your own repo. Take the same task, run it with a raw prompt, then with a skill, and see what happens. My guess is you'll delete a lot of files from your prompt library afterward — I know I did.