The best Claude skills to install first, from my own run counts

Lists of the best Claude skills usually rank them by GitHub stars, which measures popularity rather than use. So I went through my own Claude Code logs, counted every Skill tool run since the first logged one on 28 August 2026, and picked the five I'd install first.

Below you'll find what a skill actually is, five reasons it beats pasting the same instructions into every chat, and the countdown with how to install each one. Every run count is from when I recorded this in October 2026, and each one is a floor: old logs get pruned, so the real numbers are higher.

What a Claude skill actually is

Open a skill up and it's a folder with one file inside called SKILL.md. The top of that file holds a name and a short description of when to use it. Underneath sit the instructions. Anthropic's own advice is that writing them is like putting together an onboarding guide for a new hire.

The folder can also hold reference files and scripts Claude can run. Skills work in Claude Code and in the Claude app on Free, Pro, Max, Team and Enterprise plans. In the app they need code execution switched on. On Free, Pro and Max that's the "Code execution and file creation" switch under Settings > Capabilities. Your own skills upload as a ZIP under Customize > Skills.

Why a skill beats pasting the same prompt

Say you like Claude to write your client emails a certain way, with your tone, the words you never use and how you sign off. You could paste those rules into every chat. A skill does better for five reasons.

1. 1You write it once. Pasted rules go into every new chat, and the day you leave a line out, the email comes out different. My own writing rules run to 771 words, and I'm not pasting that before every email.
2. 2Only the name and description load when a chat starts. Claude knows the email skill exists without carrying the whole thing. When I recorded this in October 2026, my folder held 37 skills whose descriptions came to 2,696 words, while the full SKILL.md files ran to 55,333. Custom instructions and project files work differently: they load whether you need them or not.
3. 3It opens itself. Ask for a client email and Claude matches it to the description and reads the full instructions, so you needn't remember the skill is there. In Claude Code you can also call it directly by typing a slash and its name.
4. 4Extra files and scripts stay shut until needed. Your email skill could keep a file of emails you were happy with, and Claude only opens it when it's writing one. Scripts are plain code, so they run the same way every time, which suits a job like checking a draft for words that make it read as AI.
5. 5You fix it once. When Claude gets something wrong, you correct the skill and every email after that gets the fix. My Upwork proposal skill keeps a dated copy before each rework: 29 June, 24 July, 28 July, 25 August, 21 September and 24 September 2026. That's at least six reworks since the end of June, and every proposal since has used the newer version.

How I ranked them, and where my ranking stops

When I recorded this in October 2026, the counts came from Skill tool calls across 839 log files in 22 project folders. To count your own, Claude Code keeps its logs as .jsonl files under ~/.claude/projects, and each Skill tool call is one run. At that point my three most-run skills were:

• unlazy: 109 runs across 100 sessions
• humanizer (my own version): 90 runs across 81 sessions
• upwork-apply, my Upwork proposal skill: 76 runs, all between 16 September and 16 October 2026

Below those, my ranking stops being useful to you. upwork-apply is private. My fourth at the time, an Excalidraw diagram skill at 25 runs, lists a GitHub page that returned a 404 when I checked on 16 October 2026, so you can't install it from there. The next three are private session tools on 15, 11 and 8 runs.

So places 1 and 2 below are ranked by my runs, and place 3 stands in for my Upwork skill. Places 4 and 5 are my picks for you, not ranked by my runs. When I recorded this in October 2026, frontend-design had 2 runs in my logs and the spreadsheet skill had 1.

The five best Claude skills to install first

5. Anthropic's document skills

These cover Excel, Word, PowerPoint and PDF, and they sit behind Claude's document features, so you may already have them without knowing. The spreadsheet one shows why they matter: its instructions say to use formulas, never hardcoded results, so the sheet still recalculates when you change an input. They're source-available, not open source.

In the Claude app they come on with file creation. In Claude Code they install from Anthropic's skills repo on GitHub, which had 181,466 stars on 16 October 2026. Run these two commands:

1. 1/plugin marketplace add anthropics/skills
2. 2/plugin install document-skills@anthropic-agent-skills

4. frontend-design

This is for anyone who's asked Claude for a website and got the site everyone else gets. It names the looks AI design clusters around, such as "a warm cream background (near #F4F1EA) with a high-contrast serif display and a terracotta or warm-clay accent", or content chopped into identical rounded cards.

Then it works in two passes: a design plan covering colour, type and layout for your brand, then a review of that plan against the brief that revises anything reading like the generic default, before any code. I built it into the skill I use to make CRMs for clients, which keeps each client's own brand colours. After the marketplace step above, it installs with /plugin install example-skills@anthropic-agent-skills.

3. skill-creator

skill-creator is the skill that builds skills. Tell it about a job you keep doing by hand and it drafts a skill, writes a few test prompts, runs them, then rewrites the skill from what you thought of the results. It comes in the same example-skills plugin as frontend-design.

It holds this slot because my own third most-run skill is one you can't install. upwork-apply takes a job post, picks my closest past projects and writes me a script for a short Loom video. Once I've recorded the Loom, it writes the cover letter to match what I actually showed, even where I drifted from its plan. It ran 76 times in the month before I recorded this. My logs don't say how I built it, so I won't claim skill-creator made it. Either way, the most useful skill you'll have is one built around your own work, because nobody else has it.

2. humanizer

Claude's writing is easy to spot. The popular free humanizer is blader's on GitHub, with 57,407 stars on 16 October 2026. It's built on Wikipedia's Signs of AI writing guide and checks for 26 patterns, rewriting each so the text reads like a person wrote it. The first pattern on its list is the "Not X but Y" sentence.

I run my own version, a separate file from blader's. It bans 34 words Claude leans on and fires by itself before any copy goes to a real person. When I recorded this in October 2026 it had 90 runs, second only to unlazy.

1. unlazy

unlazy is top, with 109 runs when I recorded this in October 2026. If Claude has ever told you a job is done when it clearly isn't, this is the fix. Before Claude starts, it writes down what done means as a list of gates, one observable outcome each, and every gate that can be run gets a check and an expected result.

A box only gets ticked once its check passes, and the proof is saved: the resolved shell, working directory, exit status, match result and an output fingerprint. The optional Stop hook is my favourite part. If Claude tries to stop while any gate is unmet, the hook blocks it and sends it back to work.

It's free on GitHub (Leonxlnx/unlazy, 3,955 stars on 16 October 2026). I've set mine to run by default on every substantial job, with the Stop hook running on every stop, and I've made some minor tweaks. My free guide to making Claude prove it's done walks through that setup. For the bigger version of this problem, when agents hand work to each other, see why a 10-agent fleet only gets you to 60 percent.

What an unused skill costs you

Every installed skill's description loads at the start of every chat. Descriptions are short next to the instructions behind them, as my own folder shows, but a skill you never run still adds its share. Install the few you'll use, count your runs after a month, and remove whatever never fires.