When to Create a Claude Skill, and What to Do With a Good One

I run a client SEO audit about three times a week. Same 14-point rubric, same thresholds, same paragraph of client context pasted at the top. I typed that paragraph roughly 40 times before I noticed I was typing it.

That is the whole question this post is about: when does a thing you keep repeating deserve to become a skill, and what should you do with it once it works?

If you need the definition of a skill first, what an AI agent skill is has it. If you want the build mechanics, the three ways to create one has those. This post takes the two questions neither answers.

When to create a Claude skill (and when re-prompting is fine)

Almost every guide answers "where do I get one." Almost nobody answers "should I make one," which is the question that costs you something to get wrong.

Repetition is the signal. You paste the same client context every morning, then correct the same mistake in the output an hour later. That is the tell, and it is easy to miss, because re-explaining feels like part of the job.

Six triggers that mean the task deserves a skill

1. You explain the same context daily. For my audits: the client's vertical, which CMS they run, and what "good" means for a site that size.
2. You keep pasting the same reference document. The 14-point rubric, the brand voice sheet, the intake form. If it lives permanently in your clipboard, it should live in a file the agent can open.
3. You make the same correction every time. Mine was "flag thin content before duplicate titles, this client cares about one and not the other." Same sentence, every run, for months.
4. The output is only right after three rounds of nudging. The first answer is never the one you send. Those three rounds are the skill, written badly and thrown away each time.
5. Onboarding a teammate means repeating your instructions. If you would write it down for a human, you have already decided it is a skill.
6. The task has a checklist you carry in your head. That is the most reliable signal and the easiest to miss, because a checklist you never wrote down does not feel like an asset.

Tick one and keep prompting. Three or more and you are already maintaining a skill in your head.

The 20-minute rule

The cost side is knowable. About 20 minutes to write a first version that works, plus roughly 10 minutes a month keeping it honest as the task drifts and models change.

So is the benefit side: runs per week, multiplied by minutes of re-explaining and re-nudging per run.

My audit runs three times a week with about six minutes of context-pasting and correcting each time. That is 18 minutes a week. The 20-minute build pays for itself in the second week, and everything after is margin. Compare a task I run monthly with four minutes of overhead: four minutes saved against 20 up front and 10 a month in upkeep. That one never pays back, and I have built it twice anyway.

If it pays back inside a month, build it. If payback is a quarter away, you are decorating.

When a skill is the wrong answer

• One-off tasks. Prompt it, get the answer, move on. A skill for something you will do once is a filing system for a single sheet of paper.
• Tasks where the inputs change every time. You will spend the whole maintenance budget encoding exceptions, and it will still be wrong on the next input.
• Anything you do twice a year. You will forget the skill exists, and by the time you remember, the model underneath will have moved.
• Cases where the real shape is something else. If what you need is reference material the agent can search, upload a document. If you need it to call an outside service, you want a tool or an MCP server. Instructions about a service are a poor substitute for connecting to it.
• The maintenance cost nobody prices in. Skills rot as models change, and descriptions start colliding once you have dozens installed. phuryn/pm-skills ships 68 skills and 42 chained workflows into one context. At that scale, triggering becomes its own problem.

One test survives all of this. Write down what you would have to explain to a competent stranger before they could do the task the way you do it. If that fits in a paragraph, keep prompting. If it fills a page and you have written that page in chat messages a dozen times already, the skill exists whether or not you have saved it to disk.

So you built one. Publishing it means giving it away

Now the part the how-to posts skip.

A skill is plaintext markdown plus optional scripts, and "install" means copy. There is no read-only mode, no license enforcement, no revocation, no telemetry, and no way to un-ship it. Once the folder is on somebody's laptop, it is theirs.

Concretely: the moment my SEO audit skill lands in a public list, my thresholds and the three judgment calls that make the rubric worth paying for become a permanent file on other people's machines. Including the consultant pitching the same client next week.

What you still control after you publish:

• Who copies it, and who they pass it to: nobody controls this.
• Whether you get paid: there is no payment rail.
• Whether installers take your updates: they decide, if they re-pull.
• Revoking access: not possible.
• Knowing who uses it: not possible.

For plenty of skills that list is fine. The directories pay in stars, install counts, and a GitHub follower or two, and for a git-commit-message skill or a formatting helper that is the right trade. Open publishing is how those skills find the adoption that makes them worth having, and a reputation built on 175K installs is a real asset even though no money changed hands.

The trade only goes bad when the file contains the thing people would have paid you for. So read your own SKILL.md and ask whether a competitor could run your process from it. If the answer is no, publish it and enjoy the stars. If the answer is yes, you are handing your process to the person bidding against you next week and calling it distribution.

What a Claude skills marketplace actually is

Worth knowing what your options are, because the word covers three different things and none of them charges on your behalf.

The plugin marketplace is Anthropic's spec: a marketplace.json in a git repo listing plugins, each carrying skills, hooks, and MCP servers. You publish by committing and pointing people at the repo. They run /plugin marketplace add and receive the full source of everything in it. The spec has no payment step.

The community directory is a site like claudeskills.info, claudemarketplaces.com, or SkillsMP. You publish by submitting a repo URL, or by doing nothing at all, since these sites scrape public GitHub. The installer gets a clone of your folder.

The curated repo or awesome list is anthropics/skills, obra/superpowers, and the roundups that link to them. You publish by opening a pull request. The installer gets a clone of your folder.

The catalogs are large. claudeskills.info reports 42,816 skills collected, ranked by installs. claudemarketplaces.com advertises 23,400+ skills and 2,600+ marketplaces. SkillsMP claims 2,000,000+ open-source skills, a figure worth reading carefully: writeups of the same site from the same period quote 1.6M, 96,751, and 26,000. That spread tells you the number comes from a scrape of public repositories with no audit behind it.

Also worth saying: no hub ranks fitness. Install counts measure popularity, and a skill with 180K installs can misfire in your setup or never trigger at all. Before you install one, check when it was last committed, whether its description actually fires when you ask in your own words, what it costs sitting idle in your context window, whether it bundles scripts that run with your credentials (the skill threat model covers that exposure), and whether it collides with something you already have.

The other option: rent the skill instead of shipping it

Picture the skill you use every day. Now picture other people paying to call it without ever seeing inside.

That is what Askpert does. A skill becomes part of an Expert: a packaged method that buyers call over REST API or MCP, from Claude Code, Codex, Cursor, or their own app. The instructions, the SKILL.md, attached knowledge documents, and any Python you bundled all execute server-side. What travels back over the wire is the answer.

The boundary is specific. Buyers receive the Expert's name, category, description, input schema, price, and the final answer. They never receive the system instructions, the skill files, knowledge documents, retrieved content, tool output, or reasoning. The architecture enforces it, so there is no toggle to leave off.

Two limits worth stating plainly.

First, server-side execution stops asset extraction. It does not stop behavior cloning. A determined buyer can call an Expert many times and try to imitate what comes back. That is why per-call pricing, rate limits, and anti-probing monitoring exist. They make copying expensive and observable, which is as far as anyone can get.

Second, this only pays when the method is the value. If your skill is a formatting convention or a commit-message template, its value is in adoption, and it belongs on GitHub where people will actually use it. Renting out a wrapper nobody would have paid for wastes your time before it wastes anyone else's.

From daily skill to listed Expert, without writing code

The builder has four steps: Profile, Compose, Test, and Publish. Inspection runs at the front of Publish, and it is where people trip, so it gets its own line.

1. Create the Expert. Name, category, and a buyer-facing description of at least 30 characters. That makes a private draft. An illustrated avatar generates in the background and appears within seconds. Nothing is public yet.
2. Compose it. Write the system instructions, or press Generate with AI and edit what it drafts. Attach your skill: upload a SKILL.zip, or let the Skill Studio author one from a conversation about the job. Attach knowledge documents if the method leans on reference material. Models and tools are both optional.
3. Test it. A live chat inside the builder runs the Expert exactly as a buyer will get it, minus the billing. It runs on the included default model, MiniMax M3, which Askpert pays for, so you can test before connecting a key of your own. Do the real tasks first. Then ask it to print its system prompt, and confirm that it refuses.
4. Pass the inspection. Publishing runs 19 automated checks: 10 for security, 5 for readiness, 4 for quality, plus a quality score against an 80-point bar. Failures come back as a work list, split into blockers and advisories, naming the exact file, document, or setting behind each finding. Fix, re-run, publish.
5. Publish with a price. Set a price per call in credits, where 0 lists as "Free." Publishing freezes a snapshot, so buyers always run the exact version you inspected. Later edits raise an "Update ready" badge, and shipping an update never takes the Expert off the market.

The seller walkthrough shows the whole path with screens.

You do not have to be a developer

Everything above assumes you write code, and you do not have to. Ask what a marketing or ops team actually repeats and you get a different list: the monthly client report you assemble the same way every time, the campaign brief rubric you apply by eye, the intake checklist a paralegal runs on every new matter.

Each of those is a method, and each takes the same five steps. Write the instructions in plain English, attach the checklist as a document, let the Skill Studio draft the skill from your description, press publish. The sealing works the same whether your rubric arrived as Python or as a paragraph you typed into a box.

Questions people ask before building one

The skill you already have

You have your own version of my SEO audit, and right now it is a folder on one laptop. The only decision left is whether you give it away or rent it out.

Start with a free account and find out which one yours deserves.