Every list of the best Claude skills is a list of coding skills. Test runners, debugging playbooks, a thing that scrapes websites. Useful if you ship software. Most people I know who use Claude every day do not ship software. They run meetings, chase decisions, write the email nobody wants to write, and try to get a spreadsheet to admit what it knows.
I catalogued 141 skills for THE INDEX and wrote the authoring standard behind the free library. Along the way I installed far more than I kept. This is the set that survived a year of real work, what each one replaced in my week, and the six I pulled out and why.
If you only want the files: every skill named below is in the free library, no signup, one SKILL.md each. Paste it into Claude.ai, Cowork or Claude Code and it works.
What Claude skills are, in one paragraph
A skill is a folder with a file called SKILL.md at the top. The file opens with a short header: a name and a description. Claude keeps the name and description of every installed skill in view, and when a request matches one, it opens that file and follows the steps inside. Nothing else loads until it is needed, which is why you can have forty skills installed without paying for forty skills on every message.
They work in Claude.ai, in Cowork, in Claude Code and through the API. The format is an open standard, so one folder behaves the same way everywhere. Anthropic shipped skills in late 2025 and publishes its own set on GitHub. You can write one in a text editor in ten minutes.
That is the whole mechanism. Marketplaces, plugins and the thousands of community skills are packaging around that one idea.
Claude skills vs projects vs agents vs prompts
People mix these up because they overlap. One table, then I will stop.
| What it is | Loads when | Use it for | |
|---|---|---|---|
| Prompt | Text you type | Every time you type it | One-off requests |
| Project instructions | A standing brief attached to a project | Every message in that project | Things that are always true: who you are, what the company sells, house style |
| Skill | A folder with a SKILL.md | Only when a request matches its description | A repeatable job with steps: turn notes into actions, run a pre-mortem |
| Agent / subagent | A separate Claude with its own context window | When the main conversation hands work to it | Heavy or parallel work you want kept out of your main thread |
The line I use: project instructions describe you, a skill describes a job. If you have pasted the same five-step procedure into a chat three times, that is a skill. If you keep re-explaining what your company does, that is project instructions.
The 14 Claude skills I actually use
Grouped by what they do to my week. Each name is the folder name in the free library, so you can search it directly.
Writing
plain-language-rewrite. Turns dense professional prose into something a smart outsider can read once and understand. I run it on anything going to a board, a client or someone outside the team. It replaced the second draft I never had time for.
cut-by-thirty. Removes a third of the words and keeps every claim that carries weight. It is strict about leaving numbers, names and commitments alone. It replaced "make this shorter", which always cut the wrong things.
opening-line-fixer. Takes a slow first paragraph and gives the reader a reason to keep going by the end of the first sentence. I use it on cold emails and on blog posts, this one included. It replaced staring at the first line.
one-pager-compress. Any long document into one page that stands on its own. The one page is what actually gets read. It replaced sending a 20-page deck and hoping.
Meetings and comms
meeting-notes-to-actions. A transcript or a wall of notes becomes three lists: decisions, actions with an owner and a date, open questions. Its best habit is refusing to invent an owner nobody named. Unowned and undated items go to the top, because those are the ones that will not happen. It replaced minutes nobody read.
bad-news-email. Puts the bad thing in the first two lines instead of burying it under context. It replaced the cowardly draft I would have written myself.
status-update. Leads with what changed and what the reader has to do. It replaced the weekly update that was really a diary.
Decisions
premortem. Assume the plan has already failed, list the causes, fix what you can while there is still time. Ten minutes before kick-off, every time. It replaced the retro where we said "we should have seen that."
steelman. Builds the strongest case against your position before you commit to it. Uncomfortable, and worth it. It replaced asking a colleague to poke holes, which they were always too polite to do properly.
Sales and outreach
cold-open. An opening line that earns the second line. That is the whole spec, and the skill holds itself to it. It replaced "I hope this finds you well."
objection-preempt. Handles the objection you know is coming, on your terms, before it is raised. It replaced the follow-up call where the objection had already set.
Working with Claude itself
skill-description-fixer. Rewrites a skill's description so it fires at the right moment and only then. Descriptions written as topic labels do not fire. Descriptions written as situations do. If you install one meta-skill, make it this one. It replaced wondering why a skill never triggered.
hallucination-trap. Builds checks into a workflow that catch invented detail before it reaches a reader. I run it on anything with numbers or names in it. It replaced finding the made-up statistic after the email went out.
skill-creator (Anthropic, official). Scaffolds a new skill and tests whether its description triggers. This is the one official skill on every install I run. Anthropic's document skills for Word, Excel, PowerPoint and PDF are also worth keeping if you produce files.
The six I deleted, and why
- Persona skills. "Act as a senior copywriter with 20 years' experience." The whole file is a character sheet. One sentence in project instructions does the same job, and it does not compete for attention with skills that contain an actual procedure.
- Skills with a label for a description. "Helps with emails." It never fired, because Claude matches on situations, not nouns. You can fix it with skill-description-fixer, but in my experience an author who got the description wrong usually got the body wrong too.
- The 900-line skill. Everything about SEO in one file. Every time it fired it flooded the context and pushed out the details of what I had actually asked for. A skill body should fit on a phone screen. Heavy reference belongs in a separate file the skill points to.
- The duplicate of a default. "Summarise this document." Claude summarises documents. A skill that restates the model's default behaviour costs a slot and adds nothing.
- Two skills that fought. I had two email skills with overlapping descriptions and neither had an exclusion. Which one fired was a coin toss. I kept one and gave it a "not for" line.
- A skill with a script I could not read. Skills can run code. If I cannot read the script in two minutes, it does not go on my machine. Same rule as any other software.
How to tell a good Claude skill in ten seconds
Open the SKILL.md and read the description. Four checks:
- It starts with "Use when" and names a situation, not a topic.
- It includes the words a real person would type, including the sloppy ones.
- It has a "not for" clause pointing at its nearest neighbour.
- Further down, the body has a gotchas section and a way to tell whether the output is right.
Here is the header from skill-description-fixer, which passes its own test:
---
name: skill-description-fixer
description: Use when an agent skill exists but never triggers, or triggers on the wrong requests, because its description is written as a label rather than an activation condition. Not for writing the skill body.
---Compare that with "Skill for improving skill descriptions." Same topic, and it would never fire on a real request.
Best Claude skills by job
For marketing: headline-ladder gives you a graded set of headlines so you choose your risk level. proof-point-picker selects the evidence that answers a specific audience's specific doubt. analogy-finder finds an analogy that carries the actual mechanism instead of a vibe. tone-shift moves a draft between registers without touching the facts.
For writing: the four above, plus read-aloud-audit for sentences that fall apart when spoken and transition-repair for paragraphs that stopped following each other.
For sales: cold-open and objection-preempt, then discovery-questions, follow-up-sequence, lost-deal-postmortem and pricing-explainer. That set covers a deal from first line to post-mortem.
For data: csv-profiler before you draw a conclusion, numbers-sanity before you send one, delta-explainer when someone asks why the number moved.
For coding: the developer lists cover this ground well and I will not repeat them. From Anthropic's own set I keep skill-creator and frontend-design. From the library: commit-message, pr-description, error-triage and rollback-plan.
How to install a Claude skill
Claude.ai and Cowork: zip the skill folder and upload it under Skills in your settings.
Claude Code: put the folder in .claude/skills/ inside the project, or in ~/.claude/skills/ to have it everywhere.
API and Agent SDK: attach the skill to the request and it loads the same way.
Then test it. Type a request in your own words and see whether the skill fires. If it does not, the description is the problem, not the body.
Questions people ask
Are Claude skills free?
The format is open. Anthropic's skills are free on GitHub. The 108 in the free library are free to use, modify and redistribute, with no attribution required.
Do skills work in Claude.ai or only in Claude Code?
Both, and in Cowork and through the API. Claude Code has the richest tooling around them, but the file itself is identical.
How many skills should I install?
As many as have a distinct trigger. In practice that is ten to twenty. Past that, descriptions start to collide and the wrong one fires.
Can I use a Claude skill in ChatGPT or Gemini?
The file is plain markdown. Paste it into a Project or a Gem as instructions and most of the procedure carries. What you lose is the automatic triggering, which is a Claude feature. THE INDEX ships each skill with install notes for Copilot, Gemini Gems and Grok for this reason.
What is the difference between a skill and a custom instruction?
An instruction is always on and describes you. A skill is on demand and describes a job.
Where to get them
All 108 single-purpose skills are in the Free Skill Library, downloadable one at a time or as a zip. The larger catalogue, with call numbers, packs and per-tool install paths, is THE INDEX. Anthropic's official skills are on GitHub under anthropics/skills.
And if you would rather someone set the whole thing up for your actual job, instructions and task files included, that is what The Pack is. Paste your LinkedIn, pay $49, get the files in 24 hours.
Leave a comment: