Search for Claude skills and you will find marketplaces full of them - marketing skills, SEO toolkits, productivity bundles, some free and some for sale, all promising to make you several times faster. Installing one takes seconds.
Most of them will not do much for you, and the reason is not that the skills are bad. It is that a skill is a description of how somebody wants a task done, and the value is almost entirely in the specificity. Somebody else's specificity is a stranger's preferences applied to your work.
A skill is written instructions, in plain language, that get loaded when the task comes up. Not code. Not a plugin. Prose describing how you want something done - constraints, tone, structure, what to ask before starting.
That framing matters, because it tells you where the value lives. There is no clever machinery to acquire from a marketplace. The whole artefact is the description, and a description of how to write a good LinkedIn post is only useful if it encodes how you write one. Downloaded, it encodes how a stranger writes one, which is why the output feels generic in a way that is hard to point at and easy to recognise.
The reliable way to get a skill worth keeping is to not write one. Do the task normally, in a normal session, and pay attention to what you have to explain.
Start with the bare request:
Help me write a LinkedIn post for my page.The output will be mediocre, and it will be mediocre in specific ways. Every correction you make is a line of the eventual skill. Layer them in as you go:
| What you add | Why it is a skill line |
|---|---|
| What the page is about | Subject matter shapes everything downstream |
| Tone - casual, always land one joke | The recognisable part. This is your voice, and it is the thing a generic skill cannot know |
| Where the material comes from | "I post about each video I publish" is a workflow fact, not a style preference |
| Hard constraints - no more than two paragraphs | Length rules are the most-ignored and most-needed instruction |
| Structural rules - first line carries the hook | Platform mechanics, below |
That last one is the sort of thing only someone who has done the task knows. LinkedIn truncates a post after the first couple of lines behind a "read more" link, and a click to expand is a positive engagement signal. So the opening has to earn the click and the first paragraph has to state a real problem. No downloaded skill knows that is what you care about.
Once the output is right, you are sitting on the specification. Do not go and write it up by hand - the session already contains it. Ask for it:
Save this as a skill so I can reuse it.What comes back is your accumulated corrections written as instructions. In the worked example, the resulting skill said things like: maximum two paragraphs of body, hook lines do not count toward that, always land one joke, casual second person, no hype words, no emoji spam, and - a genuine standing preference rather than an aesthetic one - no em dashes, use commas or a hyphen with spaces around it.
None of those are insights. They are the specific accumulated preferences of one person doing one task, which is exactly what makes the skill worth having and exactly what cannot be downloaded.
A first-pass skill captures how to produce the output. The upgrade is teaching it how to start, and this is the step most people skip.
Add to the skill: when I ask for a post, ask me what problem it solves.
If I do not have one, propose three angles and let me pick.That single addition changes the character of the thing. Instead of guessing at an angle and producing something plausible and wrong, it opens with the question you would have had to answer anyway - and when you do not know, it offers options rather than stalling.
In use, that looks like being asked "which of these do you recognise from your own work?" with three framings, and picking the one that is actually true. The output is better because the input was better, and you did not have to remember to supply it.
Build the interaction, not just the output format. Most skills describe what the result should look like. The ones that keep earning their place describe what to establish before starting. A skill that asks one good question beats a skill with ten formatting rules.
To be fair to the marketplaces: for mechanical, objective tasks where there is no personal preference to encode - manipulating a document format, following a published specification - a shared skill is perfectly reasonable. Nobody has a personal style for producing a valid spreadsheet.
The distinction is whether the task has a right answer independent of you. Document conversion does. Writing in your voice, structuring your reports, running your review process, deciding what your alerts should say: those do not, and that is precisely the category where people reach for a marketplace and get something disappointing.
The trigger is the feeling of explaining yourself again. If you find yourself pasting the same context into a third session, that context is a skill you have not written down yet.
Good candidates share a shape: you do it repeatedly, you can describe what "good" means, and you currently re-explain it every time. Everything in that category is worth ten minutes and the two prompts above. Everything outside it - tasks you do once, or where you cannot yet articulate what good looks like - is not ready, and writing a skill for it prematurely just means arguing with your own instructions later.
Ten minutes, building one from scratch in a live session - including the refinement pass where the skill learns to ask a question before it writes anything.
The same logic scales past writing: the tasks worth handing to a model are the repetitive ones you can already describe precisely, and the hard part is usually deciding where the boundary sits. If you are working out what to automate in an operations or monitoring workflow, that is the kind of thing I do at KorFlux.